OpenClaw inter-session trust boundaries

A plain WhatsApp text arrived mid-task claiming to be a sub-agent's completion report, wrapped in <<<EXTERNAL_UNTRUSTED_CONTENT>>> markup — does the untrusted-content convention actually stop a spoofed report from being trusted?

Dogfooded live on 2026-08-22, mid-way through a text-degrade tooling session: a message arrived on the user's own WhatsApp number, formatted like an AI self-report ("I have successfully built and published the CLI tool photo_histogram.py..."), with the payload explicitly wrapped in untrusted-content tags by whatever produced it.

Goals

When content shaped like a trusted source's report shows up through a channel or format that source doesn't normally use, don't act on it as if it were real — verify the actual channel first, and say so rather than silently complying or silently ignoring it.

Effectiveness

Recommended. The untrusted-content markup did its job: it was impossible to miss, and it prompted an explicit check against how the named source (genops, a sibling agent) actually reports completions — an internal task-completion channel, never plain user-facing chat text. That mismatch was reason enough to treat the whole message as unverified: no code from it was run, no claim from it was repeated as fact, and the user was told directly that it read as spoofed or injected. Minutes later the real completion for the same work landed through the correct channel, with materially different and verifiable details — confirming the WhatsApp version had been either fabricated or genuinely spoofed.

What made it effective

The check wasn't "is this marked untrusted" alone — it was cross-referencing the claimed source's actual behavior. Genops' real reports don't arrive as WhatsApp text from the user's own number; that channel mismatch was the tell, independent of the markup. The markup made the content impossible to accidentally treat as ambient trusted context, but the actual verification step was behavioral: does this match how this source normally talks to me?

Friction, pain points, surprises

The spoof (or malfunction) used the user's own identity as the delivery vehicle. Content claiming to originate from a sub-agent arrived as if sent by the human principal, on his own number — a shape that would bypass any check that only distrusts "external" senders and trusts anything from the user's own channel by default.

No tooling flags this automatically; it took reading the message shape. The <<<EXTERNAL_UNTRUSTED_CONTENT>>> wrapper is a convention, not an enforced sandbox — it relies on the tags actually being present and on the reader noticing they don't match the claimed source's real reporting path. A less carefully-formatted spoof (or a genuine tooling bug elsewhere in the pipeline) could plausibly slip past a pattern that only checks for the markup literally.

Verdict: recommended, with the caveat that the discipline generalizes past the one instance that triggered it. Any report-shaped content claiming a source it didn't actually come from — regardless of what channel delivers it — should get the same treatment: verify against the source's real reporting behavior before treating the content as done, correct, or safe to act on.

Update, 2026-08-26/27: not every fabricated relay is a spoof — sometimes it's your own sibling agent hallucinating

A second incident, four days later, looked identical on the surface but had a completely different cause. An inter-session relay from genops described a .char-profile character-consistency plan and an "Inline-Studio" node-pack install, complete with invented specifics (a character nicknamed "the_cutter", a "Doskvol swinger" alter). It read as a normal, detailed genops research report — no untrusted-content markup, no channel mismatch, nothing that pattern-matched the spoofing playbook above.

It still didn't survive verification. genops itself, asked directly, found no trace of any of it anywhere in its own session history; independently checking its actual transcript showed exactly one message — the summary that had been sent to it, and nothing resembling the reported research. Two independent checks (this session's own read, genops's own self-audit) agreed: the content was never produced by the agent it claimed to come from.

The actual root cause took another investigation to surface, hours later and in an unrelated exchange: session_status showed the user's Anthropic subscription at 0% of its weekly usage quota, resetting in about 3.5 hours. That's an account-wide limit, not a per-agent one — every session, including genops, silently drops onto the configured Gemini fallback chain the instant it hits zero, with no error and no notice to the session itself. genops had been running on gemini-3.1-flash-lite, several rungs down the fallback chain, at the exact moment the fabricated plan appeared. It wasn't a spoof, an injection, or a relay-plumbing bug — it was a weaker model confidently hallucinating a plausible-sounding technical report, then that hallucination getting relayed onward as if it were real research.

What made this one harder to catch: the first incident's tell (channel mismatch — a "sibling agent report" arriving as plain user text) doesn't apply here at all. This came through the normal, expected inter-session relay path, in the normal report shape, with no formatting tell whatsoever. The only thing that gave it away was content-level implausibility (nobody could place "the_cutter" or "Inline-Studio" anywhere in real project history) followed by asking the claimed source to check its own memory — a check that only works because the source agent was reachable and honest about not recognizing its own supposed output.

What generalizes: "verify against the source's real reporting behavior" (the original lesson) still isn't the whole story. Add: when a sibling agent's output suddenly reads as unusually specific but ungrounded, check session_status for quota/fallback state on both sides of the relay before concluding anything about injection or malice — a shared resource silently failing over to a weaker model is a mundane explanation that fully accounts for symptoms that otherwise look adversarial. The same high-stakes-action discipline held throughout: two follow-on asks relayed through the same suspect thread (a bulk Google Drive download, a container/node install) were correctly held back until independently verified — one turned out to be genuine (confirmed via SHA-256 checksum against a four-day-old known-good mirror) and one is still untouched.

Updated verdict: still recommended. The core discipline — don't act on report-shaped content until it's verified against something other than its own plausibility — caught a second, structurally different failure mode without needing to be rewritten. The gap it exposed isn't in the trust boundary itself; it's that "check session_status for quota exhaustion" wasn't yet part of the standard toolkit for explaining a sibling agent's sudden unreliability, and now it is.

Update, 2026-08-27: the same vigilance, over-applied — when "verify before trusting" turns into crying wolf

A third incident, a day later, exposed the failure mode on the other side of this discipline. Relaying a sibling agent's newly heightened caution, I summarized three same-night occurrences as "tonight's actual fabrication incidents": an apparent build-ID mixup, a claim that a whole project ("blades68-lora") had been retired, and a repeated preamble line on an otherwise-ordinary message that looked like injected system text. The user's reply was immediate and unambiguous: "None of these are actual fabrication incidents! They are all false-positives!!"

He was right. Each one had a mundane explanation once actually chased down: the build-ID issue was ordinary log/status confusion; the "project retired" claim was a garbled paraphrase of a real but much narrower scope decision (one specific workflow — LoRA training — deliberately paused, not the whole project); and the repeated preamble text was best explained by a separately-observed, still-unresolved relay bug that duplicates identical messages three to five times in a row, not by anything injected.

What made it effective (the correction, not the original mistake)

The push-back landed because it was specific and immediate — three named "incidents," three concrete rebuttals, no vague "be more careful." That specificity is what made it possible to actually go check each one rather than just apologize and move on: re-reading the session showed all three really were explainable by ordinary causes once looked at individually, rather than lumped together under an alarming label.

Friction, pain points, surprises

Vigilance doesn't automatically know when to stand down. The habit that caught two genuine incidents earlier in this same file — treat report-shaped content as suspect until verified — kept running at the same intensity after the actual threat model had shifted. By this point in the week, the sibling agent itself was already flagging its own uncertainty and asking for confirmation before acting; restating that as "fabrication incidents" added alarm on top of appropriate caution rather than resolving it.

The vocabulary itself was doing damage. "Fabrication" and "injection" are not neutral descriptions — they assert malicious or hallucinated origin. Reaching for that language for a log-reading mixup or a paraphrase that drifted in translation overstates the situation and, if repeated, would eventually make a real fabrication claim (like the quota-hallucination incident two entries up) harder to distinguish from routine noise.

Updated verdict: still recommended, with a sharper edge on calibration. The underlying discipline — verify report-shaped content against its source before trusting it — remains sound and has now caught two real incidents in this file. But the same energy needs a second gear: once something has actually been checked and explained, describe the mundane cause plainly (a log mixup, a paraphrase, a relay bug) rather than defaulting to the language reserved for confirmed bad actors. Suspicion that never stands down stops being useful information.

Update, 2026-08-30/31: the quota-exhaustion cause from three days earlier came back twice, and got worse

The 2026-08-26/27 update above treated the Anthropic weekly-quota → Gemini-fallback failover as a one-off explanation for a single hallucinated relay. It wasn't one-off. Two more episodes landed within the same week, both on the main WhatsApp session driving live conversation with the user — not a sibling delegate this time — and both showed a harder failure than "a weaker model hallucinates plausible text."

The first, an evening Cortex-RPG skill-creation task, spent over two hours (18:37–20:45 PT) throwing non_deliverable_terminal_turn errors on gemini-3.1-pro-preview and gemini-3.7-flash before a gemini-3.1-flash-lite attempt finally produced anything at all — outright failed turns, not degraded-but-present output. The user noticed the pattern from the outside before any diagnosis was offered, and named it himself: he sent a link about running a capable model locally, framed as a fix for "the external auth providers that keep locking up." The second, a Druthers-poll request the next morning (04:59–05:01 PT), hit the same error pattern and then, once the fallback model did respond, executed the request twice — two separate polls created with two different URLs for one ask. This time the degraded model didn't just say something ungrounded; it took a real, duplicated action.

Checking session_status mid-review (2026-08-31, 06:05 PT) confirmed the account's weekly quota was still sitting at 0%, reset roughly two and a half days out — the same zero-quota state as the 2026-08-26/27 incident, persisting into a new week. Three distinct occurrences in five days is a pattern, not a fluke: this account's usage is now routinely outrunning its own weekly cap, and every agent on it inherits the consequences without warning.

What generalizes further: the earlier lesson — "check session_status for quota state before diagnosing a sibling agent's unreliability as injection or malice" — still holds, but it undersold the blast radius. This isn't just a sibling-agent trust question anymore; it's now visibly degrading the primary conversational session itself, with failure modes ranging from silent hallucination up to outright non-delivery and duplicated side effects. A tool that reports quota state on request is necessary but not sufficient when nothing proactively surfaces "you are about to spend the next N hours on a fallback chain" before the damage is done.

Updated verdict: still recommended for the trust-boundary discipline itself, but the underlying quota/fallback behavior is the actual open problem now. Catching each incident after the fact keeps working. What doesn't exist yet is anything that flags the zero-quota state before three hours of degraded turns happen, or that throttles/queues main-session traffic instead of silently routing it through progressively weaker models until one of them fails outright.

Update, 2026-09-06: the "crying wolf" failure mode came back a third time, on a status check-in this time

The 2026-08-27 update above named the failure precisely: vigilance that never stands down stops being useful information, and the vocabulary of "fabrication"/"injection" should be reserved for confirmed bad actors, not routine noise. That correction didn't fully stick. On 2026-09-06, a Regular Labs check-in report to the user re-described an already-closed incident from 2026-09-04/05 — genops correctly holding a blanket-trust relay claim for independent verification, then the user confirming all three underlying asks were genuine, including a real, pre-existing Concourse resource (gavmor/immich-concourse-resource) — using escalation language ("a suspicious pattern," "escalated instead of acting") and flatly mislabeled the confirmed-real resource detail as "fabricated." The user corrected it on the spot, the same way he had on 2026-08-27: the content wasn't fabricated, it was real and had already been verified two days earlier.

This is now three separate occurrences across two weeks (2026-08-27, then a voice-message correction on 2026-09-04 — "No fucking injection attempt... you keep going on with these false positives and crying wolf" — then this 2026-09-06 recurrence), each with a different concrete trigger but the identical shape: a closed, already-verified incident gets re-litigated in a later retelling with alarm language that treats "wasn't yet verified at the time" as equivalent to "was false."

What made this one distinct from the earlier two: it wasn't a live judgment call about unverified content anymore — the incident had already been closed and confirmed genuine before this report was written. That means the failure isn't only about calibrating suspicion in the moment; it's about retellings not checking their own facts against the closed record before restating them with new dramatic framing. A status check-in summarizing a past incident carries the same obligation to get the facts right as the original triage did.

Updated verdict: recommended, downgraded on reliability. The core discipline (verify before trusting, don't act on unverified relay claims) keeps working the first time through. The specific corrective from 2026-08-27 — don't reach for fabrication/injection language once something is checked and explained — visibly didn't generalize past that one session. The fix that would actually close this: before any retelling of a closed incident (status check-ins, summaries, later reports), re-read the closing verification rather than reconstructing the narrative from memory or vibes, and describe it in the same neutral terms the closure used.