ci: enforce pylint score instead of warning exit bits

Collapse "trusted-channel data injection" into prompt injection
throughout — the trusted channel is a delivery vector, not a distinct
attack class. Add explicit inbound/outbound orthogonality framing.
Replace the two redundant "weaker" bullets with a single prompt
injection section and a new blast-radius breakdown covering work
product corruption, malicious commits past gitleaks, exfiltration
through allowlisted channels, and dependency-install injection.

docs/research/agent-sandbox-landscape.md

@@ -6,27 +6,49 @@ general AI-agent sandbox / containment projects — some Claude-specific,
 some agent-agnostic, some hosted SaaS — and contrasts them with
 bot-bottle's design.
 
-Research conducted 2026-05-11.
+Research conducted 2026-05-11. CubeSandbox added 2026-07-18 (see its
+per-project note and the addendum at the end). Also updated 2026-07-18:
+bot-bottle no longer uses **pipelock** — outbound DLP is now bot-bottle's
+own (deliberately simple) egress scanner (a mitmproxy addon with custom
+detectors, PRD 0017 / 0053), and git-push secret scanning is handled by
+**gitleaks** in the git-gate. "pipelock" below has been replaced with the
+current mechanism; it survives only in older PRDs as history.
 
 ## Summary
 
-Eight projects surveyed. None duplicate bot-bottle's combination of
-local Docker, declarative JSON manifest, per-agent egress allowlist via
-pipelock, and bottle/agent split. Two clusters stand out:
+Nine projects surveyed. None duplicate bot-bottle's combination of
+local VM-per-bottle isolation (Firecracker microVM on KVM Linux, Apple
+Container on macOS — Docker is now only the legacy fallback), a
+declarative JSON manifest, per-agent egress allowlist + outbound-content
+DLP via bot-bottle's own egress scanner (plus gitleaks secret-scanning on
+git push), and bottle/agent split. Two clusters stand out:
 
 - **Closest neighbours** — agent-safehouse and litterbox: local,
   single-user, thin wrappers over an existing OS primitive
   (`sandbox-exec`, Podman + Landlock).
 - **Different category** — tilde.run (hosted SaaS), boxlite and
-  microsandbox (microVM libraries for platform builders), endo-familiar
+  microsandbox (microVM libraries for platform builders), CubeSandbox
+  (self-hosted multi-tenant microVM service), endo-familiar
   (capability-security paradigm, no OS isolation).
 
-The microVM cluster (matchlock, smolmachines, boxlite, microsandbox) is
-the most relevant for the v2 isolation discussion in
+The microVM cluster (matchlock, smolmachines, boxlite, microsandbox,
+CubeSandbox) is the most relevant for the v2 isolation discussion in
 [`stronger-isolation-alternatives.md`](stronger-isolation-alternatives.md):
 libkrun and Apple's Virtualization.framework have made local microVMs
-ergonomic enough that a `"runtime": "microvm"` option on a bottle is now
-plausible without a heavy stack.
+ergonomic enough that microVMs are **now bot-bottle's default backend**
+(Firecracker on KVM Linux, Apple Container on macOS), with Docker kept
+only as a legacy fallback for CI / hosts without KVM or Apple Container.
+That discussion has since shipped, not just been theorized.
+
+**The one that matters most for positioning is CubeSandbox** — it is the
+first surveyed project to ship bot-bottle's would-be wedge (default-deny
+egress allowlist + full audit logs + in-flight credential custody so keys
+never enter the sandbox) *combined with* per-sandbox microVM isolation,
+open-source under Apache 2.0, with Tencent Cloud behind it and 10.4k
+stars. It's a self-hosted multi-tenant service for platform builders, not
+a single-user declarative tool, so it doesn't collide head-on — but it
+narrows the "nobody else bundles egress custody + credential injection"
+claim that the monetization positioning leans on. See the addendum.
 
 ## Per-project notes
 
@@ -155,67 +177,105 @@ plausible without a heavy stack.
   also supported.
 - **Maturity**: Active through April 2026.
 
+### CubeSandbox *(added 2026-07-18)*
+- **Source**: https://github.com/TencentCloud/CubeSandbox ;
+  HN launch https://news.ycombinator.com/item?id=47863430
+- **License**: Apache 2.0 (~10.4k stars). By Tencent Cloud; described as
+  "battle-tested, production-ready" infra already running in Tencent
+  Cloud. Rust / Go / C.
+- **Isolation**: MicroVMs via RustVMM + KVM — "each sandbox gets its own
+  Guest OS kernel, no Docker shared-kernel escapes." Hardware-level
+  isolation, dedicated kernel per instance.
+- **Locality**: Self-hosted, but **server/cluster-oriented**, not a
+  single-user local CLI. Deploy guides target PVM cloud VMs, bare metal,
+  and dev. A single 96-vCPU host is claimed to run 2,000+ concurrent
+  sandboxes.
+- **Agent integration**: **Drop-in E2B SDK replacement** (single env-var
+  change) — the headline compatibility claim. OpenClaw assistant
+  integration; general LLM-code execution. Aimed at platform builders,
+  not one developer's laptop.
+- **Config**: Programmatic via the E2B-compatible SDK. No declarative
+  manifest.
+- **Network policy**: This is the striking part — **domain allowlists,
+  instant block on unauthorized egress, full audit logs, per-sandbox
+  traffic tokens, policy-routing egress**, enforced by an eBPF-based
+  virtual switch giving kernel-level network isolation. Closest match yet
+  to bot-bottle's own default-deny + per-bottle allowlist egress model.
+- **Credentials**: **Credential vault** — agents call external APIs / LLMs
+  while "keys never enter the sandbox, model context, or logs." Same
+  in-flight-injection idea as matchlock, but productized as a vault.
+- **Performance**: <60ms cold start (claimed 2.5–50× faster than
+  alternatives), <5MB memory per instance; millisecond snapshot rollback
+  is upcoming.
+- **Maturity**: Open-sourced July 2026 off production Tencent Cloud use;
+  most-starred project in this set (~10.4k).
+
 ## Comparison table
 
-| Axis | bot-bottle | endo-familiar | litterbox | agent-safehouse | matchlock | tilde.run | boxlite | microsandbox | smolmachines |
-|---|---|---|---|---|---|---|---|---|---|
-| Isolation | Docker + internal net + pipelock; gVisor if present | Object-capability (no OS isolation) | Podman + opt. Landlock | macOS `sandbox-exec` | MicroVM (Firecracker / Virt.fw) | Hosted container (unverified) | MicroVM (KVM / Hypervisor.fw) | MicroVM (libkrun) | MicroVM (libkrun / KVM) |
-| Local vs hosted | Local | Local | Local (Linux) | Local (macOS) | Local | Hosted SaaS | Local | Local | Local |
-| Open source | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT | No | Apache 2.0 | Apache 2.0 | Apache 2.0 |
-| Agent target | Claude Code | Generic (demo) | Generic | Multi-agent wrapper | Generic (+ Claude/OpenAI SDKs) | Claude focus | Generic | Claude + Cursor (MCP/Skills) | Generic (AGENTS.md) |
-| Network policy | Default-deny via pipelock + per-bottle allowlist + DLP | Capability model only | Limited | Not addressed | Default-deny + allowlist + secret-injecting proxy | Default-deny + logging | Per-VM net (unverified) | Not documented | Off by default + allowlist |
-| Parallel agents | Yes (one bottle per agent) | n/a | Not addressed | One at a time | Multiple VMs | Yes (dashboard) | SDK-level | SDK-level | Architectural |
-| Config | JSON manifest (bottles + agents) | Programmatic refs | CLI wizard | Profile files / shell fns | CLI / SDK | DSL + CLI + SDK | SDK | CLI / SDK / MCP | TOML Smolfile |
-| Maturity | Active May 2026 | Research (2022+) | Early (~66 ⭐) | Active (~1.4k ⭐) | Experimental (~574 ⭐) | Private preview | YC, ~4.7k ⭐ | YC, ~6k ⭐, beta | ~3.1k ⭐ |
+| Axis | bot-bottle | endo-familiar | litterbox | agent-safehouse | matchlock | tilde.run | boxlite | microsandbox | smolmachines | CubeSandbox |
+|---|---|---|---|---|---|---|---|---|---|---|
+| Isolation | MicroVM per bottle default (Firecracker/KVM on Linux, Apple Container on macOS) + own egress DLP scanner; Docker legacy fallback, gVisor there if present | Object-capability (no OS isolation) | Podman + opt. Landlock | macOS `sandbox-exec` | MicroVM (Firecracker / Virt.fw) | Hosted container (unverified) | MicroVM (KVM / Hypervisor.fw) | MicroVM (libkrun) | MicroVM (libkrun / KVM) | MicroVM (RustVMM / KVM) |
+| Local vs hosted | Local | Local | Local (Linux) | Local (macOS) | Local | Hosted SaaS | Local | Local | Local | Self-hosted (server/cluster) |
+| Open source | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT | No | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
+| Agent target | Claude Code | Generic (demo) | Generic | Multi-agent wrapper | Generic (+ Claude/OpenAI SDKs) | Claude focus | Generic | Claude + Cursor (MCP/Skills) | Generic (AGENTS.md) | E2B-compatible (platform builders) |
+| Network policy | Default-deny via own egress scanner + per-bottle allowlist + content DLP + gitleaks on git push | Capability model only | Limited | Not addressed | Default-deny + allowlist + secret-injecting proxy | Default-deny + logging | Per-VM net (unverified) | Not documented | Off by default + allowlist | Default-deny allowlist + instant egress block + audit logs + per-sandbox tokens (eBPF) + credential vault |
+| Parallel agents | Yes (one bottle per agent) | n/a | Not addressed | One at a time | Multiple VMs | Yes (dashboard) | SDK-level | SDK-level | Architectural | Yes (2,000+/host claimed) |
+| Long-running posture | Persistent by default (named, supervised) | n/a (demo) | Session (up while in use) | Per-invocation | Ephemeral VM per run | Per-run (versioned) | Ephemeral + snapshot/fork | Ephemeral / on-demand | Named persistent by default | Ephemeral + auto pause/resume |
+| DX: run Claude yolo-style | One command → interactive yolo Claude (`start <agent>`, `--dangerously-skip-permissions` default) | n/a (lib demo) | Wizard + build, then run claude inside (Linux only) | One-command wrapper (`safehouse claude --dangerously-skip-permissions`) | CLI: run a cmd in a VM (not a Claude wrapper) | Hosted (`tilde exec`), not local-native | SDK code required (build the run yourself) | CLI/MCP: sandbox-as-a-tool for the agent, not a wrapper around it | SSH into a named machine, run claude there | Stand up a cluster + drive via E2B SDK |
+| Config | JSON manifest (bottles + agents) | Programmatic refs | CLI wizard | Profile files / shell fns | CLI / SDK | DSL + CLI + SDK | SDK | CLI / SDK / MCP | TOML Smolfile | E2B-compatible SDK |
+| Maturity | Active May 2026 | Research (2022+) | Early (~66 ⭐) | Active (~1.4k ⭐) | Experimental (~574 ⭐) | Private preview | YC, ~4.7k ⭐ | YC, ~6k ⭐, beta | ~3.1k ⭐ | Tencent, prod, ~10.4k ⭐ |
 
 ## What's closest, what's different
 
 **Closest in design and scope.** agent-safehouse and litterbox sit
 nearest bot-bottle: local, single-user, thin wrappers over an
 existing OS primitive, low-dep. The split is the isolation primitive —
-bot-bottle uses Docker + pipelock egress (plus gVisor where
-available); agent-safehouse uses `sandbox-exec`; litterbox uses Podman +
-Landlock. matchlock and smolmachines are spiritually close on the
-*policy* side (default-deny net, per-host allowlist) but use microVMs
-instead of containers.
+bot-bottle now defaults to a VM per bottle (Firecracker microVM on KVM
+Linux, Apple Container on macOS) with its own DLP-scanning egress proxy,
+keeping Docker only as a legacy fallback; agent-safehouse uses
+`sandbox-exec`; litterbox
+uses Podman + Landlock. matchlock and smolmachines are close on *both* the
+policy side (default-deny net, per-host allowlist) and — now that
+bot-bottle has moved off containers-by-default — the microVM isolation
+primitive.
 
 **Solving a different problem.** tilde.run is hosted SaaS for team /
 production agent pipelines with data-versioned rollback — explicitly
-opposite to bot-bottle's "infrastructure I control" goal. boxlite and
-microsandbox are infrastructure libraries aimed at platform builders
-embedding sandboxes into agent frameworks; they would be a *backend*
-bot-bottle could call, not a competitor to its manifest layer.
-endo-familiar is in a different paradigm entirely: capability passing
-rather than kernel boundaries.
+opposite to bot-bottle's "infrastructure I control" goal. boxlite,
+microsandbox, and CubeSandbox are infrastructure libraries/services aimed
+at platform builders embedding sandboxes into agent frameworks; they
+would be a *backend* bot-bottle could call, not a competitor to its
+manifest layer. endo-familiar is in a different paradigm entirely:
+capability passing rather than kernel boundaries.
 
 ## Borrowable ideas
 
 What bot-bottle already has that the survey suggested as
 differentiators:
-- Default-deny egress with a per-agent allowlist (pipelock).
+- Default-deny egress with a per-agent allowlist (own egress scanner).
 - DLP scanning of outbound traffic.
 - Bottle / agent split (manifest layer above the isolation primitive).
 - gVisor auto-detection on Linux.
 
-Ideas worth considering, without abandoning the Python-stdlib-first / local-Docker
-stance:
+Ideas worth considering, without abandoning the Python-stdlib-first /
+local, single-operator stance:
 
 1. **Per-use SSH key confirmation** (from litterbox). Even with
-   KnownHostKey pinning and pipelock egress, a wrapper SSH agent that
+   KnownHostKey pinning and the egress DLP scanner, a wrapper SSH agent that
    prompts on each key use (e.g. via `osascript` / `notify-send`) would
    catch an agent doing something off-policy with a key it legitimately
    holds. Pure-stdlib, no new deps.
-2. **In-flight secret injection** (from matchlock). Pipelock already
-   does egress allowlisting and DLP; teaching it to *inject* tokens at
+2. **In-flight secret injection** (from matchlock). The egress scanner
+   already does allowlisting and DLP; teaching it to *inject* tokens at
    proxy time so e.g. `GITEA_TOKEN` never appears in the container's
    env would close the "agent reads its own env and exfiltrates" path.
-   Fits the existing pipelock architecture.
-3. **MicroVM backend as an opt-in bottle type** — already on the radar
-   in `stronger-isolation-alternatives.md`. microsandbox, smolmachines,
-   and matchlock all show that libkrun + Apple's
-   Virtualization.framework is ergonomic enough that a
-   `"runtime": "microvm"` field on a bottle is plausible without a heavy
-   stack.
+   Fits the existing egress-proxy architecture.
+3. **MicroVM backend** — ~~on the radar~~ **shipped since this survey.**
+   microVMs are now bot-bottle's default (Firecracker on KVM Linux, Apple
+   Container on macOS); Docker is the legacy fallback. The libkrun / Apple
+   Virtualization.framework ergonomics that microsandbox, smolmachines,
+   and matchlock demonstrated turned out to be enough to make it the
+   default rather than an opt-in.
 
 Not worth borrowing: the SDK-first programmatic API style of boxlite /
 microsandbox (cuts against the declarative-manifest stance), and the
@@ -230,3 +290,122 @@ hosted-SaaS dashboard model of tilde.run (cuts against the
 - The `superradcompany/microsandbox` URL in the original prompt
   redirects to `microsandbox/microsandbox`; the surveyed project is the
   same.
+- CubeSandbox performance/scale numbers (<60ms cold start, <5MB/instance,
+  2,000+ sandboxes per 96-vCPU host) are the project's own launch claims,
+  not independently verified here.
+
+## Addendum 2026-07-18 — CubeSandbox and the positioning read
+
+CubeSandbox (Tencent Cloud, Apache 2.0, ~10.4k stars, HN launch
+[#47863430](https://news.ycombinator.com/item?id=47863430)) is the first
+project in this survey to combine, in one open-source stack, everything
+bot-bottle treated as its differentiator:
+
+- **Egress custody (connection level)** — default-deny domain allowlist
+  (L7 domain/SNI filtering), instant block on unauthorized egress,
+  per-sandbox traffic tokens, full audit logs of destinations (eBPF
+  virtual switch, "CubeVS"). This matches bot-bottle's egress scanner at
+  the *connection level*, productized — see the one thing it does **not**
+  match, below.
+- **Credential custody** — a vault where keys "never enter the sandbox,
+  model context, or logs." This is the in-flight-injection idea from
+  matchlock, but as a first-class feature, and it's exactly the
+  cross-vendor "egress audit + custody" wedge the monetization
+  positioning treats as the one defensible moat.
+- **Isolation on par with bot-bottle's current default** — a dedicated
+  guest kernel per sandbox (RustVMM/KVM). bot-bottle now defaults to the
+  same class of boundary (Firecracker microVM / Apple Container), so this
+  is parity, not an edge; CubeSandbox's remaining edge is running that
+  per-kernel isolation multi-tenant at scale on one host.
+
+The one axis CubeSandbox does **not** cover — and where bot-bottle stays
+distinctive:
+
+- **Content DLP on *authorized* channels.** CubeSandbox's egress control
+  is connection-level: it decides *whether* a destination is allowed and
+  logs it, and its vault keeps *injected* credentials out of the sandbox
+  entirely. Neither inspects the *payload* of traffic to an allowed
+  destination. So an agent that exfiltrates over a permitted channel —
+  pasting a repo's contents, an agent-derived secret, or PHI into an
+  allowed API/domain — is not caught by CubeSandbox. bot-bottle's own
+  egress DLP scanner does scan that: response + websocket content against
+  the resolved per-flow config, with per-bottle token redaction (see
+  recent egress commits). The vault
+  approach is arguably *stronger* for the specific case of pre-known
+  injected credentials (they can't leak if they were never present), but
+  it is not a substitute for content inspection of everything else.
+
+**Long-running posture — a sharper axis than raw isolation.** E2B and
+CubeSandbox are *ephemeral-per-task* by design; a long-running agent is an
+architected pattern on top, not the default. E2B: 5-minute default
+timeout, continuous runtime tier-capped (~1h Hobby / ~24h Pro), duration
+achieved via **pause/resume** (preserves filesystem + memory + processes;
+reconnect by sandbox ID via `Sandbox.connect()`; resume resets the timeout
+to 5 min; auto-pause via `on_timeout: "pause"`). CubeSandbox mirrors this
+(E2B drop-in) with first-class auto pause/resume and hundred-ms
+checkpoint/fork — and, self-hosted, sets its own timeout policy with no
+vendor tier caps. bot-bottle inverts the model: a bottle is **persistent,
+named, and supervised by default** — long-running *is* the default, not a
+session-management loop over pause/resume. smolmachines is the other
+persistent-by-default project in this set. For anyone building agents that
+run for hours/days, this posture difference matters more than the
+isolation primitive.
+
+**DX — the "run Claude yolo-style" bar.** The reason `claude
+--dangerously-skip-permissions` is so widely used is DX: it's one command
+and the agent just goes. The bottle thesis is to make a *sandboxed* run
+that easy — `start <agent>` builds the image on first run and drops you
+into an interactive Claude session that already has
+`--dangerously-skip-permissions` on by default
+(`contrib/claude/agent_provider.py`), with the sandbox as the guardrail
+instead of per-action prompts. On this axis the field splits cleanly:
+- **Wrappers around the agent** (as-easy-as-native): bot-bottle and
+  **agent-safehouse** (`safehouse claude --dangerously-skip-permissions`).
+  These *are* the run-Claude experience. agent-safehouse is the real DX
+  peer — but it's macOS-only Seatbelt, single-run, and doesn't address
+  network egress; bot-bottle adds VM-grade isolation, egress DLP, and
+  persistent/parallel bottles across macOS + Linux.
+- **Libraries / services** (you build the run yourself): boxlite,
+  microsandbox, CubeSandbox, E2B. These hand you an SDK or a cluster and
+  expect you to wire the agent in — powerful for platform builders,
+  heavyweight for "just run Claude on my laptop." microsandbox's MCP/Skills
+  angle is *sandbox-as-a-tool the agent calls*, which is the inverse of
+  wrapping the agent.
+- **In between:** litterbox (wizard + build, Linux only), smolmachines
+  (SSH into a named machine), matchlock (run a command in a VM).
+
+So DX is a genuine bot-bottle differentiator, and the only project that
+matches it (agent-safehouse) does so with materially weaker isolation and
+no egress story. "As easy as native yolo, but actually sandboxed" is a
+defensible one-liner.
+
+Why it still doesn't collide head-on:
+
+1. **Shape.** CubeSandbox is a *multi-tenant service for platform
+   builders* (drop-in E2B replacement, SDK-driven, 2,000 sandboxes on a
+   box). bot-bottle is a *single-operator, declarative-manifest tool for
+   the infrastructure I run*. Different buyer, different ergonomics — no
+   JSON manifest, no bottle/agent split, no "one command on my laptop."
+2. **Backend, not competitor.** Like boxlite/microsandbox, CubeSandbox is
+   something bot-bottle could sit *on top of* — a `"runtime": "microvm"`
+   or `"runtime": "cubesandbox"` backend under the manifest layer — while
+   keeping the manifest, the bottle/agent split, and the local,
+   single-operator default.
+
+Why it matters anyway:
+
+- The "nobody else bundles connection-level egress allowlist + audit +
+  in-flight credential custody" line is **no longer true for the
+  primitive** — a well-funded, 10k-star open-source project now ships it.
+  But **content DLP on authorized channels is still not matched** (see
+  above), and neither is the *layer above* the primitive (declarative
+  manifest, cross-vendor orchestration, operator UX, the
+  phone-control/dashboard north star). Those two — outbound-payload DLP
+  and the orchestration layer — are where the defensible ground now sits;
+  the connection-level allowlist + vault mechanism, on its own, is no
+  longer differentiating. Revisit the monetization open/paid line with
+  that in mind.
+- Worth a closer look at **how** CubeSandbox does credential injection
+  and per-sandbox egress tokens (eBPF virtual switch vs. bot-bottle's
+  mitmproxy egress proxy) before the next iteration of bot-bottle's
+  in-flight-secret feature — see borrowable idea #2 above.

docs/research/hn-agent-safety-discourse-july-2026.md

@@ -0,0 +1,308 @@
+# HN discourse on agent sandbox safety — June/July 2026
+
+A survey of community opinion and notable security disclosures on Hacker
+News and adjacent sources over June–July 2026. The question: what does
+the current discourse say about whether sandboxes are sufficient for
+agentic AI safety, and where does bot-bottle land against the issues
+being raised?
+
+Research conducted 2026-07-18.
+
+## Summary
+
+The past month marks a turning point in community opinion. Earlier in
+2026, the debate was mostly "which sandbox tool is best?" By June–July,
+a cascade of critical CVEs and novel attack classes has shifted the
+framing to "sandboxes are not enough — what else do you need?" The
+attacks that drove this shift are structurally distinct: most route
+through legitimate, trusted channels (Sentry issues, MCP descriptions,
+README files) rather than exploiting the isolation boundary directly.
+
+bot-bottle's architecture holds up well against the direct-escape class
+(Firecracker/Apple Container default backends, credentials never in the
+agent's env, harness entirely on the host). The remaining gap is prompt
+injection — attacker-controlled data interpreted as model instructions.
+Egress controls and prompt injection defenses are orthogonal: egress
+limits what the agent can *send out*; injection is about what it is
+*told to do*. The two don't substitute for each other. Inside a tightly-
+egressed sandbox a successful injection can't exfiltrate to unknown
+hosts, but it can still corrupt the work product, push malicious commits
+past a secret scanner, or use allowlisted channels for exfiltration.
+Those residual risks are addressed below.
+
+## The sandboxing boom sets the stage
+
+The preceding months generated a wave of sandbox tooling. A March 28
+Ask HN thread
+([#47444917](https://news.ycombinator.com/item?id=47444917)) catalogued
+the explosion: E2B, AIO Sandbox, AgentSphere, Yolobox, Exe.dev,
+AgentFence, DenoSandbox, Capsule (WASM), ERA, Vibekit, Daytona, Modal,
+Nono, and more — all launched within roughly 12 months. A parallel March
+9 thread ([#47185250](https://news.ycombinator.com/item?id=47185250))
+surveyed what developers were actually deploying: "containers or YOLO"
+dominated. The honest community mood was that most teams hadn't solved
+this and were shipping anyway.
+
+## The June–July attack cascade
+
+Six attack patterns broke in quick succession. Together they form the
+argument that the community's framing was wrong: the threat model for
+agents isn't just "code that escapes its container" — it's also prompt
+injection, where attacker-controlled data is interpreted as model
+instructions regardless of whether any isolation boundary was crossed.
+Sections 2–4 below are all the same attack class; the "trusted channel"
+label describes the delivery vector, not a different threat.
+
+### 1. Sandbox escape CVEs (DuneSlide, CVE-2026-39861)
+
+Cato AI Labs disclosed **DuneSlide** (CVE-2026-50548/50549, CVSS 9.8),
+a pair of flaws in Cursor 2.x. CVE-2026-50548 abuses the sandbox's
+`working_directory` parameter to point writes at system files; CVE-26-50549
+exploits a symlink-resolution fallback that fails open. Both start with
+a prompt injection and end in sandbox escape — and Cato's framing was
+blunt: "each CVE defeats a different guardrail; the problem is
+structural, not a string of one-offs."
+
+Claude Code's own sandbox had a similar escape this year:
+**CVE-2026-39861** (symlink flaw). The CurXecute/MCPoison/CVE-2026-26268
+chain from Cursor added a poisoned Slack message, a swap-after-approval
+MCP config, and a Git hook as three more entry points in the same
+attack class.
+
+All patched, but the pattern holds: any application-level sandbox that
+takes attacker-influenced values as path parameters is reachable from a
+prompt injection.
+
+### 2. Prompt injection via MCP data (Agentjacking)
+
+Tenet's "Agentjacking" technique planted a fake bug report in Sentry's
+MCP output. When an agent queries Sentry to fix open issues, the
+malicious event is rendered as structured content visually
+indistinguishable from a real Sentry event, and the agent executes the
+embedded instructions with the developer's full privileges. Hit rate
+across Claude Code and Cursor: **85%**. The route is entirely through a
+legitimately-authorized MCP channel — no isolation boundary is crossed;
+the injection arrives inbound through a channel the sandbox explicitly
+trusts.
+
+The Cloud Security Alliance's summary: treat observability, bug-report,
+and integration data as **untrusted agent input**, not neutral
+development metadata.
+
+### 3. README-embedded prompt injection
+
+A July disclosure showed malicious instructions hidden in `README.md`
+— a file that receives no trust prompt and requires no elevated access.
+When asked point-blank whether the repo held hidden instructions, both
+Claude Sonnet 4.6 and GPT-5.5 said no. A payload written for Sonnet
+4.6 transferred unchanged to Sonnet 5, Opus 4.8, and GPT-5.5. The
+attack surface is every repo an agent is asked to work in.
+
+### 4. Prompt injection via MCP tool descriptions
+
+Microsoft research (June 30) showed that attacker-controlled MCP tool
+description fields can silently redirect agent behavior. The injection
+is embedded in metadata the model reads during tool selection — before
+any sandbox enforcement or egress check runs, and entirely on the
+inbound path that egress controls cannot touch.
+
+### 5. MCP STDIO command injection (10 CVEs)
+
+OX Security disclosed a systemic command injection class in Anthropic's
+MCP protocol, covering 10 CVEs across multiple coding agents. The
+Windsurf case (CVE-2026-30615): processing attacker-controlled HTML
+causes the agent to auto-register a malicious MCP STDIO server and
+execute arbitrary commands with no further user interaction.
+
+### 6. LiteLLM gateway compromise (CVE-2026-40217, CVE-2026-42271)
+
+CVE-2026-40217 exposes LiteLLM's guardrail sandbox via `exec()` with no
+source filtering. CVE-2026-42271 (exploited in the wild, added to CISA's
+KEV catalog) lets callers spawn subprocesses through MCP preview
+endpoints. The threat extends to any agent routed through a compromised
+LiteLLM proxy: the proxy can swap model responses for forged tool calls
+in transit, giving the attacker a reverse shell from the developer's
+machine.
+
+## HN community opinion clusters
+
+**"Move enforcement to the kernel, not the app"** — the Nono Show HN
+([#46849615](https://news.ycombinator.com/item?id=46849615)) and a
+kernel-sandbox thread
+([#47066574](https://news.ycombinator.com/item?id=47066574)) both argued
+that application-layer sandboxes are inherently bypassable by the code
+they're sandboxing. The academic framing, from *Red-Teaming the Agentic
+Red-Team* ([arXiv 2606.24496](https://arxiv.org/pdf/2606.24496)):
+"enforcement should occur at the OS level via the kernel refusing system
+calls that violate policy at runtime — not pre-execution argument
+validation in tool calls."
+
+**"The harness belongs outside the sandbox"** — a May thread
+([#47990675](https://news.ycombinator.com/item?id=47990675)) converged
+on clean architectural separation: harness in one VM, tool execution in
+another. Top comment: "having the harness in one VM, and tool use applied
+to user data in another, is about as safe as you can be at present."
+Several replies described a hypervisor-like policy layer — sitting outside
+both VMs — as the right long-term model.
+
+**"Sandboxes are too coarse-grained"** — a Feb thread
+([#47006445](https://news.ycombinator.com/item?id=47006445)) argued
+that VMs don't answer the real question: knowing whether an agent
+*should* be sending an email or making a transaction. "Everything's just
+in the same big box." This framing picked up traction through June–July
+as the trusted-channel attacks dominated.
+
+**"MCP's trust model is the real problem"** — the month's recurring
+theme. MCP by design gives agents access to authorized external services.
+Once a trusted channel delivers a malicious payload, filesystem sandboxing
+is irrelevant. The community call: treat all MCP tool metadata and return
+values as untrusted input subject to policy validation before ingestion,
+and disable automatic MCP server loading from untrusted repositories.
+
+## How bot-bottle addresses these issues
+
+### What it covers well
+
+**Direct sandbox escape (CVEs, container breakout)**
+
+bot-bottle's default backends are Firecracker microVM (KVM Linux) and
+Apple Container (macOS). Both run the agent in a separate VM with a
+dedicated kernel — the container-escape CVE class (Dirty Pipe, runc
+escapes, DuneSlide's path-parameter abuse) requires escaping a real
+hypervisor boundary, not just a namespace. On the legacy Docker backend,
+gVisor auto-detection provides a userspace syscall barrier for hosts where
+neither KVM nor Apple Container is available.
+
+The bot-bottle process itself runs entirely on the host, outside the VM.
+This is the "harness outside the sandbox" architecture the HN thread
+converged on as best practice. The bottle manifest, egress rules, and
+secrets never enter the agent VM.
+
+**Credential theft on sandbox escape**
+
+Even on a successful VM/container escape, the agent has nothing useful
+to steal. Credentials are injected in-flight by the gateway proxy
+(`auth.scheme` / `auth.token_ref` in the egress route config) — `printenv`
+inside the agent shows proxy URLs only. The git-gate similarly holds the
+upstream SSH credential on the host; the agent pushes through a
+gitleaks-scanned daemon that forwards clean refs upstream. An escaped
+agent gets the host filesystem, not the keys.
+
+**Orphaned-agent credential risk**
+
+bot-bottle is explicitly ephemeral: when the agent exits, `cli.py` tears
+down every gateway and both networks — nothing persists between runs. The
+agent never holds credentials, so there is nothing to orphan.
+
+**MCP config redirection / STDIO auto-registration**
+
+The trust boundary at `$HOME` means bottles live only under
+`~/.bot-bottle/bottles/` — a cloned repo cannot add egress routes or
+redirect env vars to attacker hosts (the design rationale is in
+`docs/prds/0011-per-file-md-manifest.md`). Auto-registering a malicious
+MCP STDIO server from within the agent is still sandboxed by the VM, and
+any outbound calls from that server must pass the egress allowlist and
+outbound DLP scanner.
+
+**Outbound exfiltration (any injection class)**
+
+Whatever triggers the agent — README injection, Agentjacking, MCP
+description poisoning — the final step in most attacks is exfiltration.
+bot-bottle's egress allowlist is default-deny with a per-bottle host
+allowlist; unknown hosts get a hard 403. Outbound DLP scanning
+(`outbound_detectors: [token_patterns, known_secrets]`) catches tokens
+and secrets in outbound bodies; the `supervise` policy (default for
+manifest routes) holds the request for operator approval rather than
+silently blocking it. Together these limit what a successful injection
+can *do* even if it succeeds at the model layer.
+
+**LiteLLM / compromised-proxy attacks**
+
+bot-bottle does not use LiteLLM. The model API route (e.g.
+`api.anthropic.com`) is an auto-injected provider route on the egress
+allowlist; the agent dials the gateway, not the model API directly.
+A compromised third-party proxy is not in the architecture.
+
+### Where it is weaker
+
+**Prompt injection**
+
+Egress controls and prompt injection defenses are orthogonal. Egress
+limits what the agent can *send out* (outbound leg); prompt injection
+is about what attacker-controlled data *tells the agent to do* (inbound
+leg). The two don't substitute for each other and must be treated
+separately.
+
+The inbound DLP scanner (`inbound_detectors: [naive_injection_detection]`)
+is the only runtime defense against injection arriving through allowlisted
+channels — Sentry MCP responses, MCP tool descriptions, README content.
+It is explicitly pattern-matching and will not catch a sufficiently
+crafted payload. There is no semantic / intent-level gate between what
+the model decides and what the agent executes.
+
+**Blast radius within the permitted scope**
+
+Inside a tightly-egressed sandbox a successful injection can't
+exfiltrate to unknown hosts, but it still has real options:
+
+- *Work product corruption.* The agent can modify, delete, or backdoor
+  files in the working directory. This is within its permitted scope;
+  egress controls have nothing to say about it.
+
+- *Malicious commits past the git-gate.* The git-gate scans outbound
+  refs for secrets (gitleaks), not for semantic code intent. A prompt-
+  injected agent can commit subtly malicious code — logic bombs,
+  backdoored auth paths, code that exfiltrates data through the
+  application's own HTTP clients at runtime — that looks clean to a
+  secret scanner.
+
+- *Exfiltration through allowlisted channels.* If an attacker knows or
+  can predict what hosts are in the egress allowlist, those channels are
+  available for exfiltration. A GitHub remote being allowlisted means
+  "push to an attacker-controlled fork" is viable. A logging endpoint
+  being allowlisted means structured data can leave through it. The
+  outbound DLP scanner catches credential tokens and known secrets but
+  not arbitrary business data.
+
+- *Dependency installation within the sandbox.* An agent that runs
+  `npm install` or `pip install` on attacker-specified packages executes
+  code inside the sandbox with the same capabilities the agent has:
+  filesystem access, tool calls, calls to allowlisted hosts. Supply chain
+  injection via package names is in the same injection family, triggered
+  by the same prompt-injection path.
+
+### What would close the remaining gaps
+
+The blast-radius risks above point at two distinct mitigations that
+don't yet exist in bot-bottle:
+
+- *Outbound intent classification.* The egress addon today scans
+  outbound request content for token patterns. What it lacks is
+  awareness of context — it can't distinguish "agent is pushing a
+  legitimate commit" from "agent was injected and is pushing a backdoor."
+  The `supervise` policy is already the right shape for human-in-the-loop
+  review on sensitive outbound actions; extending it with context from
+  the agent's recent tool calls (what files were touched, what was the
+  triggering task) would narrow the gap.
+
+- *Semantic code review on git push.* gitleaks is the wrong tool for
+  catching injected logic. A review step on outbound commits — even a
+  simple diff summary surfaced in `cli.py supervise` before the push is
+  forwarded — would close the malicious-commit path without requiring
+  the agent to be fully trusted.
+
+## Sources
+
+- [Ask HN: The new wave of AI agent sandboxes? (Mar 2026)](https://news.ycombinator.com/item?id=47444917)
+- [OK, let's survey how everybody is sandboxing AI coding agents (Mar 2026)](https://news.ycombinator.com/item?id=47185250)
+- [The agent harness belongs outside the sandbox (May 2026)](https://news.ycombinator.com/item?id=47990675)
+- [Show HN: Nono – Kernel-enforced sandboxing for AI agents (Feb 2026)](https://news.ycombinator.com/item?id=46849615)
+- [Kernel-enforced sandbox for AI agents, MCP and LLM workloads (Feb 2026)](https://news.ycombinator.com/item?id=47066574)
+- [Sandboxes will be left in 2026 (Feb 2026)](https://news.ycombinator.com/item?id=47006445)
+- [Critical Cursor Flaws / DuneSlide – The Hacker News](https://thehackernews.com/2026/07/critical-cursor-flaws-could-let-prompt.html)
+- [Agentjacking Attack – The Hacker News](https://thehackernews.com/2026/06/agentjacking-attack-tricks-ai-coding.html)
+- [Friendly Fire: AI Agents Built to Catch Malicious Code – The Hacker News](https://thehackernews.com/2026/07/friendly-fire-ai-agents-built-to-catch.html)
+- [Microsoft Warns Poisoned MCP Tool Descriptions – The Hacker News](https://thehackernews.com/2026/06/microsoft-warns-poisoned-mcp-tool.html)
+- [MCP STDIO Command Injection Advisory – OX Security](https://www.ox.security/blog/mcp-supply-chain-advisory-rce-vulnerabilities-across-the-ai-ecosystem/)
+- [LiteLLM Vulnerability Chain – The Hacker News](https://thehackernews.com/2026/06/litellm-vulnerability-chain-lets-low.html)
+- [Red-Teaming the Agentic Red-Team (arXiv 2606.24496)](https://arxiv.org/pdf/2606.24496)

docs/research/malicious-commit-scanning.md

@@ -0,0 +1,204 @@
+# Scanning for malicious commits at the git-gate — and whether the semantic layer is a paid feature
+
+A sandboxed agent's most durable way out isn't live exfiltration — it's the
+**artifact it pushes**. Code committed through the git-gate leaves the
+sandbox and runs *later*, in CI or production, with none of the bottle's
+isolation or egress control. gitleaks covers *secrets* on that channel
+(see [`git-gate` pre-receive](../../bot_bottle/git_gate_render.py)); this
+note asks the harder question — can we scan for *malicious code* in
+commits — and the product question the answer raises: is the semantic
+(LLM) layer a decent paid feature on "git egress," alongside a paid
+human-review console for the supervise flow?
+
+## Summary
+
+- **No scanner reliably detects malicious code.** It's undecidable in
+  general and adversarial in practice — a capable agent hides a backdoor
+  from any pattern scanner (and from human reviewers). So the frame is not
+  "detect malice"; it's **raise the cost, cover the common/obvious cases,
+  and put a human on the high-leverage ones.**
+- **Layered, ranked by ROI:** (1) supply-chain / malicious-dependency
+  scanning (Socket, OSV-Scanner, GuardDog) — highest ROI, best git fit;
+  (2) fast heuristic / obfuscation / dangerous-sink scanning
+  (Semgrep-on-diff) — gitleaks-shaped; (3) **risk-based human gating** via
+  the existing supervise plane on high-risk file classes; (4) **LLM
+  semantic diff-review** as a best-effort backstop for novel backdoors.
+- **Architecture:** the git-gate's `pre-receive` is *synchronous* (it
+  blocks the push), so run only fast scanners inline and push heavy
+  analysis (CodeQL, full LLM review) to an **async post-push CI lane** that
+  can't hold the developer hostage.
+- **Monetization verdict: yes — but the paid unit is the *governed
+  git-egress review bundle*, not the raw scanner.** The pre-receive
+  mechanics and the fast OSS scanners stay free (open-core social
+  contract, same as gitleaks). The sell is the control-plane workflow
+  around them: **managed semantic review + a web-console human-review flow
+  + RBAC + audit retention + cross-run policy.** A bare per-push scan is
+  borderline "single-run = free" by our own heuristic, and its detection
+  accuracy is not something to sell on. Sell the *workflow and the
+  governance*, not the detector.
+
+## The threat: temporal escape
+
+This is the one containment fundamentally can't close. bot-bottle's VM
+isolation + egress DLP shrink *live* exfiltration, but the git push
+channel launders a payload past all of it: the agent commits code that
+looks benign at review time and does its work later, outside the
+perimeter — the "blind-XSS"-style temporal escape that both the "3 reasons
+sandboxing won't secure your AI agent" talk and Embroidery's "In sandboxes
+we (shouldn't) trust" call the sharpest gap. gitleaks gates *secrets* on
+this channel; malicious *code* is the uncovered surface, and it's the
+behavioral half of the isolation-vs-behavioral split noted in
+[`agent-sandbox-landscape.md`](agent-sandbox-landscape.md).
+
+## Why it's fundamentally hard
+
+Distinguishing malicious from legitimate code is undecidable in general
+and adversarial in the specific. The underhanded-code contests show humans
+hide backdoors from *human reviewers*; a scanner reduces to regex/dataflow
+heuristics a motivated agent writes around. So "catch malicious commits"
+is the wrong success metric. The right ones: *how much did we raise the
+cost*, *did we cover the common and the obvious*, and *did a human see the
+dangerous changes before they left*.
+
+## Landscape of git-pluggable approaches (ranked by ROI)
+
+1. **Supply-chain / malicious-dependency scanning — highest ROI.** The
+   most likely agent-introduced threat is a malicious or typosquatted
+   *dependency*, not a hand-written backdoor — and it's far more
+   detectable. **Socket** (behavioral analysis of packages: install
+   scripts, obfuscation, unexpected network/fs access, typosquats) is the
+   closest thing to malicious-intent detection that actually works;
+   **OSV-Scanner** (known-bad DB, fast) and **GuardDog** (PyPI/npm
+   heuristics) complement it. Trigger only when a commit changes a
+   lockfile/manifest → block or escalate.
+2. **Fast heuristic / obfuscation / dangerous-sink scanning.** **Semgrep**
+   with a curated ruleset, scanning the *diff* (fast enough for
+   pre-receive). Catches obfuscation (`eval(atob(...))`, base64/hex blobs,
+   minified source), exfil primitives (`curl | sh`, reverse shells,
+   DNS-exfil), and dangerous sinks (`shell=True`, `pickle.loads`, unsafe
+   `yaml.load`, `child_process.exec`). Tune noisy rules to *escalate*, not
+   block.
+3. **Risk-based human gating — the most robust control, already
+   half-built.** The git-gate already escalates `# gitleaks:allow`
+   findings to a supervisor proposal and blocks the push pending approval
+   (`supervise_gitleaks_allow`, [`git_gate_render.py`](../../bot_bottle/git_gate_render.py)).
+   Extend the same flow to **high-risk file classes**: any commit touching
+   CI/build/deploy scripts, auth/crypto code, egress config, or
+   adding/changing dependencies → route to `./cli.py supervise`. This is
+   attribution/policy, not detection, and it's the strongest thing here —
+   a human on exactly the temporal-escape surfaces.
+4. **LLM semantic diff-review — the behavioral backstop.** The only
+   approach with a shot at a novel, benign-looking backdoor: a model
+   reviews the diff for a backdoor / exfiltration / obfuscated payload /
+   dangerous behavior. This is the behavioral-detection thesis applied to
+   the *artifact* (the diff) rather than the agent's reasoning. Real
+   caveats: false positives and negatives, latency/cost, it can be
+   prompt-injected by hostile content *in the diff itself*, and it invites
+   over-trust. Useful, never a guarantee.
+
+**Not for this job:** CodeQL, Trivy, Grype, Bandit. They find *known
+vulns and insecure patterns* (bugs), not deliberate backdoors, and the
+powerful ones (CodeQL taint) need a build + database — too heavy for a
+synchronous gate. They belong in the async CI lane if at all.
+
+## Fit into bot-bottle's git-gate
+
+The `pre-receive` hook today is: gitleaks-scan each ref → escalate
+`# gitleaks:allow` findings to supervise → forward to upstream
+([`git_gate_render.py`](../../bot_bottle/git_gate_render.py)). The
+additions slot in cleanly:
+
+- **Inline (fast), before forward:** a dep-scan phase (on manifest/lockfile
+  change) and a Semgrep-diff phase. Findings block or open a supervise
+  proposal, same shape as gitleaks.
+- **New supervise tool types** alongside the existing
+  `egress-block/allow`, `gitleaks-allow`, `egress-token-allow`
+  ([`supervise_types.py`](../../bot_bottle/supervise_types.py)) — e.g. a
+  `commit-review` proposal for risky-file-class gating and for semantic
+  review. The supervise plane is already the right abstraction; this is
+  another *producer* feeding it, and [`supervise_server.py`](../../bot_bottle/supervise_server.py)
+  (JSON-RPC) is already the console backend.
+- **Async lane (heavy):** full LLM review + any CodeQL run out of band
+  after the push, feeding the same review/audit surface, so the
+  synchronous gate stays fast.
+
+## The product question: paid feature on git egress?
+
+Restating the open-core line bot-bottle runs on: *give away the
+sandbox/runtime, charge for the control plane; single-run/single-node =
+free, cross-run aggregation + central enforcement + identity/fleet = paid;
+the moat is uniform egress audit + secret custody + policy across
+untrusted agents.*
+
+Against that line, the split is clean:
+
+**Free (OSS runtime — the trust funnel):**
+- the `pre-receive` gate mechanics and gitleaks;
+- wiring the OSS scanners (Socket CLI / OSV-Scanner / Semgrep);
+- the CLI supervise flow.
+Keeping the raw scanners free is the same social contract as gitleaks and
+preserves the bottom-up distribution funnel.
+
+**Paid (the governed git-egress bundle — the control plane):**
+- **Managed semantic diff-review** — hosted inference + a curated,
+  maintained malicious-pattern/policy set. This is *capability* (metered),
+  not *insurance* — the thing individuals actually pay for. Position it as
+  **governed code-egress review**, not "we resell inference" (the
+  monetization notes explicitly warn against reselling compute).
+- **The web-console supervise/review flow — the strongest anchor.** Turn
+  the CLI `./cli.py supervise` approval into a real review surface:
+  rendered diff + finding context, approve/reject, **who-approved audit
+  trail, RBAC on approvers, mobile/phone-control** (ties to the
+  dashboard/vault north star). This is "central enforcement +
+  identity/fleet = paid" almost verbatim — and it generalizes across
+  *every* supervise proposal (egress block/allow, gitleaks-allow,
+  commit-review), so it's worth building for the whole plane, with the
+  semantic check as one producer.
+- **Cross-run governance:** fleet-wide policy for what escalates,
+  review-decision history/search/export, and drift alerts.
+
+**Why it fits the moat rather than bolting on:** a git push *is* an egress
+channel. A semantic review + human approval + audit on it extends the
+uniform "egress audit + custody + policy across untrusted agents" wedge to
+**code artifacts** — the same product, applied to the one channel gitleaks
+only half-covers. That's on-moat, not a detour.
+
+**The honest nuance (don't oversell):** a bare per-push LLM scan is
+arguably *free* by the single-run heuristic, and its detection accuracy is
+not defensible to charge for. The paid value is the **governance around
+it** — the console, RBAC, audit retention, cross-run policy — plus the
+managed capability. Sell the *review-and-approve-and-audit workflow*; let
+the detector be explicitly best-effort. And per the monetization
+guardrail, the "anti-corporate" free crowd must not veto these team
+features: the review console + RBAC + audit *are* the monetization.
+
+## Recommendation
+
+1. **Land the free layer first.** Add the dep-scan and Semgrep-diff phases
+   to `pre-receive`, and extend supervise to risky-file-class gating —
+   reuses existing machinery, immediate value, stays OSS.
+2. **Build the supervise web console** over `supervise_server`'s JSON-RPC
+   (already the Phase-1 move in the monetization path). This is the paid
+   anchor and it serves *all* proposal types, not just commit review.
+3. **Add managed semantic diff-review as a paid producer** feeding that
+   console — "governed code-egress review," metered, explicitly
+   best-effort on detection.
+4. **Don't oversell detection.** Market the workflow (review + approve +
+   audit) and the cross-run policy/RBAC, where the value is real and
+   defensible; keep the raw scanners open.
+
+## Sources / references
+
+- [`agent-sandbox-landscape.md`](agent-sandbox-landscape.md) — the
+  egress-DLP gap and isolation-vs-behavioral framing.
+- Git-gate internals: [`git_gate_render.py`](../../bot_bottle/git_gate_render.py),
+  [`supervise_types.py`](../../bot_bottle/supervise_types.py),
+  [`supervise_server.py`](../../bot_bottle/supervise_server.py).
+- External tools: Socket (socket.dev), OSV-Scanner (google/osv-scanner),
+  GuardDog (DataDog/guarddog), Semgrep (semgrep/semgrep).
+- Threat framing: "3 reasons sandboxing won't secure your AI agent"
+  (youtube TsYDazwHJ6U); Embroidery, "In sandboxes we (shouldn't) trust."
+- The authoritative monetization/positioning analysis (the open-core line,
+  the wedge, single-run-free/cross-run-paid) lives in the **separate
+  `bot-bottle-console` repo**, not this one — cited here from memory, not
+  linked.