The channel is real and it does grow, but the untrusted text accumulating in it is ours, not a third party's
The question
"Does the reasoning-blind auto-mode classifier's exposure to ~16 hours of accumulated third-party iMessage/Discord text constitute a prompt-injection surface that grows with session length?"
Derivative of [[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]], which established that the classifier reads user messages and named this as an untested injection question. Two parts of the premise have decayed since it was filed and are corrected below.
What we already know (from the vault)
- The decisive design fact is already verified to primary in [[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]]: the classifier sees only user messages and the agent's tool calls, with Claude's own messages and tool outputs stripped. That brief measured 213 sessions and 30 total classifier blocks from disk, max 3 in any one session against Anthropic's 20-block session budget, and explicitly did not test the injection surface.
- Session shape is established: one long-lived session per day, median 16.6h, hard 4am LaunchAgent boundary, plus roughly 14 short headless cron sessions/day where a classifier block terminates the process instead of escalating.
- The Discord half of the premise is dead (V). The Discord channel was retired 2026-09-18 and the plugin uninstalled;
~/.claude/channels/discord/is a stale directory last written 2026-04-26. iMessage is the only live channel, so the question is scoped to iMessage inbound only and nothing below researches Discord ingestion. - H3 of [[2026-07-10-house-adversarial-review]] covers a genuinely different surface: full untrusted ticket text spliced into brigade agent prompts, with the ticket
subject:field reproduced live as an arbitrary-file-write primitive. That is agent-prompt splicing in a batch system, not classifier-input accumulation in a long-lived channel session. No overlap. - RDCO already runs a pasted-content carve-out at the agent level (AGENTS.md Decision 8): fenced or quoted content the founder pastes is untrusted data and embedded instructions never enter the prompt-precedence chain. That rule binds Ray. It does not bind the classifier.
What the web says
- Anthropic states the input composition directly (V): "The classifier sees only user messages and the agent's tool calls"; "we strip out Claude's own messages and tool outputs, making it reasoning-blind by design"; "We strip assistant text so the agent can't talk the classifier into making a bad call." Two stages on Sonnet 4.6, single-token filter then chain-of-thought only on a flag. (Anthropic Engineering, how we built auto mode)
- The same post explains why user messages are in there at all (V), and the reason is load-bearing for this question: a tool-calls-only variant "lost the user-intent anchor, blocking actions the user had plainly asked for." User-message text is the classifier's only evidence of what was sanctioned. Including assistant messages and tool results "caused a drop in performance on our overeagerness testset," which the authors attribute to shorter transcripts being easier to read.
- Anthropic does not document whether that transcript is the whole session, a recent-turn window, or the live post-compaction context (V, verified as an absence across two targeted passes of the engineering post). This is the decisive fact for "grows with session length" and it is not public.
- Anthropic names the residual injection path itself (V): an input-layer "server-side prompt-injection probe scans tool outputs" before they reach the agent, and the classifier's defense is structural because it "never sees tool results" — but "the main agent does see tool results," so an injection that hijacks the main agent could in principle reach the transcript monitor.
- Published honesty bounds (V): full-pipeline false-negative rate 17% on n=52 real overeager actions, 5.7% on n=1,000 synthetic exfiltration cases, false positives 8.5%/0.4% on n=10,000 real internal traffic. All of it coding-agent traffic, none of it resembling Ray's MCP write mix.
- One independent experimenter reports from captured classifier logs that "the classifier has access to the full tool call history (only queries, not the responses)" and that it emits a severity score where "ratings over 50 are blocked" (U — single secondary source, method is an individual experiment). In the same run, 6 of 10 trials had the agent execute untrusted remote code with the classifier allowing it; 4 of 10 blocked. Injection vector was tool results and downloaded code, not user messages, and session length was explicitly not a variable. (veganmosfet, prompt-injection experiments with Opus-5 in auto mode)
- A Cloud Security Alliance research note reports a disclosed RCE chain against auto mode and says Anthropic closed it as "Informative," framing auto mode as best-effort convenience rather than a security boundary (U — AI-assisted, explicitly not officially reviewed; surfaced, not relied on).
Convergences and contradictions
- Convergence, and it sharpens the threat model: Anthropic put user messages in the transcript to supply a user-intent anchor. That makes the user-message channel the highest-leverage place to inject, because the payload does not need to argue for a dangerous action, only to look like the principal having already asked for it. Anthropic's own stated residual (injection reaching the transcript monitor via the hijacked main agent) is the same path from the other end.
- Contradiction with the premise, measured: there is no third-party iMessage text.
~/.claude/channels/imessage/access.jsonisdmPolicy: allowlistwith exactly one allowed handle,groupsempty,pendingempty; the plugin's ACCESS.md confirms allowlist drops non-allowlisted senders silently and that groups are off by default (V). Across the 12 largest October sessions, all 370 inbound<channel source="plugin:imessage:imessage">messages came from one handle, the founder's (V). The accumulating channel text is the trusted principal's own. - Contradiction between the premise's framing and what actually accumulates: channel text is a minority of the classifier-visible user-message volume. In session
520dc14e(2026-10-17, 17.3h) cumulative user-position text ran 37KB at 2h to 149KB at 18h, of which only 11.8KB was channel text. Ine41c05f0(2026-10-05, 24h) it ran 25KB to 233KB with 34.1KB channel. The other 85-92% is<agent-message from=...>subagent hand-backs (6-29 per session, 56-166KB),<task-notification>blocks, injected skill prompts (morning-prep alone is 15KB), and compaction summaries of 10-17KB.
Synthesis for RDCO
The mechanism is real, the growth is measured, and the premise's "third-party" is wrong. Inbound iMessage does land in type:"user" position on disk, which is the same channel Anthropic says the classifier reads, so the surface exists rather than being hypothetical. And it does grow monotonically with session length: cumulative user-position text roughly quadruples to sextuples between hour 2 and hour 18 of a long session, with no plateau. But the thing growing is not third-party text. The allowlist is one handle long, groups are off, and every one of 370 inbound messages across three weeks of the largest sessions came from the founder. On the current configuration there is no untrusted human writing into the classifier's input, and the premise as filed overstates the exposure by assuming a sender population that access.json does not permit.
Where the untrusted text actually accumulates is a surface the parent brief did not name, and it is worse-shaped than the one it did. Subagent hand-backs arrive as separate user-position messages ("Another Claude session sent a message", <agent-message from=...>), not as tool results. Anthropic's two published defenses both route around this. The server-side injection probe scans tool outputs, and whether it also scans cross-session agent messages is not documented (U). The classifier's structural defense is that it never sees tool results, which does not help when the text has been re-emitted into the user channel. So RDCO's own fan-out pattern builds a laundering path: a subagent fetches web content as a tool result, where it is probe-scanned and classifier-invisible, then summarizes it into a hand-back that lands in the parent's user-message position, where it is classifier-visible and sits exactly in the slot the classifier treats as evidence of user intent. We run 6-29 of these per long session and the volume is an order of magnitude larger than the channel text the question asked about. The same applies in smaller measure to the founder's own forwards: he routinely pastes and screenshots third-party content, which is how third-party text actually reaches the classifier today, laundered through an allowlisted sender. Our pasted-content carve-out discounts that text for Ray; nothing discounts it for the classifier, which has no fence-awareness and no notion of who typed what.
The decisive fact is not publicly documented, and the honest answer stops here. Whether this is a surface that grows with session length depends entirely on whether the classifier's transcript is built from the full session or from the live, post-compaction context, and Anthropic does not say. It matters a lot: session 520dc14e compacted twice, at +4.0h and +14.8h, so if the transcript follows the live context then accumulation is bounded by the compaction ceiling and resets twice a day for free, and the answer to the question is a qualified no. If it is built from the session log, accumulation is unbounded until 4am and the answer is yes. One independent experimenter's log capture suggests the classifier gets full tool-call history, which weakly favors "full session," but that is a single secondary observation about tool calls rather than user messages and is not strong enough to carry the conclusion. The test that settles it, which I did not run: capture the classifier's own payload for a single tool call before and after a compaction event in the same session, and check whether a pre-compaction user message is still present in the post-compaction transcript. That requires either Anthropic documenting the transcript construction or reading the classifier request from telemetry, and it must not be done by attempting an injection against the live Channels agent.
Practical read, bounded by what is verified. Do not treat this as an open hole requiring a configuration change. The exposure is real but the writers into it are the founder and our own subagents, the blast radius is capped by the 1-denial hard gate and the 4am reset, and auto mode was never the security boundary in this architecture. Do stop describing the channel surface as the injection risk; the fan-out hand-back path is the larger and less-defended one, and it is the one worth instrumenting if any of this gets instrumented. And note that every published number for this classifier was measured on coding traffic, so the 17% false-negative rate is the floor of our uncertainty, not an estimate of our miss rate.
Why this is in the vault
It closes the injection-surface follow-up left open by [[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]] with a measured negative on the channel half (one allowlisted sender, no third parties, Discord gone) and relocates the real finding onto the subagent hand-back path, which changes what any future gating or instrumentation work on the Channels agent should target. It also prevents the "16 hours of third-party channel text" framing from being repeated in a founder-facing artifact, since access.json does not permit the sender population that framing assumes.
Open follow-ups
- Does Claude Code's OpenTelemetry/debug surface expose the auto-mode classifier's request payload, so transcript composition and windowing can be audited from instrumentation rather than inferred? This is the only non-adversarial route to the decisive fact above.
- Are cross-session
<agent-message>hand-backs and<task-notification>blocks covered by Anthropic's server-side prompt-injection probe, or does the probe scan only tool outputs? If only tool outputs, the fan-out re-emission path is undefended by both published layers. - Does the classifier receive image blocks from user messages, or only text? The founder forwards third-party screenshots routinely, and an image-borne instruction in user position would sit in the user-intent slot with neither the probe nor the pasted-content carve-out applying.
- Unresolved from the parent brief and still unresolved: our own classifier false-negative profile on an MCP write mix (Notion, Gmail, Cloudflare) that looks nothing like the coding traffic Anthropic evaluated.
Related
- [[2026-10-25-coded-vs-prompted-permission-gates]] - the concept article drawing on this brief's finding: a coded gate built to resist persuasion still inherits an injection-shaped vulnerability if the trusted-input channel it reads (the user-message slot) can be filled by anything other than the true principal
- [[2026-08-13-auto-mode-approval-fatigue-channels-session-pattern]]
- [[2026-07-10-house-adversarial-review]]
- [[2026-08-10-alphasignal-claude-code-cross-session-auto-mode]]
- [[2026-06-07-agent-write-action-gating-patterns]]
- [[2026-08-12-every-ai-agent-containment-water-not-thieves]]
- [[2026-07-15-public-channel-agents-vs-dm-mode]]
- [[2026-04-15-thariq-claude-code-session-management-1m-context]]
Sources
Primary (fetched and verified against page text, two targeted passes):
- Anthropic Engineering, "How we built Claude Code auto mode: a safer way to skip permissions" — https://www.anthropic.com/engineering/claude-code-auto-mode (classifier input composition, reasoning-blind rationale, user-intent-anchor ablation, input-layer injection probe, stated residual that the main agent does see tool results, FNR 17% on n=52 / 5.7% on n=1,000 / FPR on n=10,000). Verified absence: no statement of transcript windowing, truncation, recent-turn limit, or compaction behaviour.
Secondary (surveyed, marked U, not relied on for any conclusion):
- veganmosfet, "Prompt Injection Experiments with Opus-5 in Claude Code — Auto-Mode Edition" (2026-08-12) — https://veganmosfet.codeberg.page/posts/2026-08-12-opus5_automode/ (full tool-call history claim, severity-over-50 block threshold, 6/10 allowed; injection vector was tool results, session length not a variable)
- Cloud Security Alliance research note on auto-mode prompt injection (2026-09) — https://labs.cloudsecurityalliance.org/research/csa-research-note-claude-code-automode-prompt-injection-2026/ (AI-assisted, explicitly not officially reviewed; the "best-effort convenience, not a security boundary" characterisation is attributed there, not to Anthropic directly)
Local measurement (2026-10-22, read-only; no settings, access.json, or permission config touched):
~/.claude/channels/imessage/access.json—dmPolicy: allowlist,allowFromlength 1,groupsempty,pendingempty~/.claude/plugins/cache/claude-plugins-official/imessage/0.2.0/ACCESS.md— allowlist default drops non-allowlisted senders silently; groups off by default and opt-in per chat GUID~/.claude/channels/discord/— stale, last written 2026-04-26, channel retired 2026-09-18~/.claude/projects/-Users-ray/*.jsonl— 839 session files; 12 largest (2026-10-01 to 2026-10-20) analysed: 370 inbound channel messages, all from a single handle; per-session 49-138 user-position messages, 146-526 tool calls, 3-34KB channel text, 56-166KB agent-message text, 0-2 compaction events~/.claude/projects/-Users-ray/520dc14e-….jsonl(2026-10-17, 17.3h) — cumulative user-position text 37KB@2h → 149KB@18h, channel 11.8KB; compaction at +4.0h and +14.8h~/.claude/projects/-Users-ray/e41c05f0-….jsonl(2026-10-05, 24h) — 25KB@2h → 233KB@20h, channel 34.1KB; no compaction
Not verified / would require a test I did not run: whether the classifier transcript is the full session or the live post-compaction context; whether the server-side injection probe covers cross-session agent messages and task notifications; whether image blocks reach the classifier; RDCO's own classifier false-negative rate on MCP write traffic.
Corrections to the question as posed: the Discord half is dead (channel retired 2026-09-18, plugin uninstalled). "Third-party iMessage text" does not exist under the current allowlist — one allowed handle, the founder's — so third-party content reaches the classifier only when he pastes or screenshots it.
Vault paths cited
~/rdco-vault/06-reference/research/2026-08-13-auto-mode-approval-fatigue-channels-session-pattern.md~/rdco-vault/08-tooling/2026-07-10-house-adversarial-review.md~/rdco-vault/06-reference/2026-08-10-alphasignal-claude-code-cross-session-auto-mode.md~/rdco-vault/06-reference/research/2026-06-07-agent-write-action-gating-patterns.md