[Bug] Subagent generated permission-bypass payload in its own assistant turn (0 tool calls) — same signature as #74223, #75616

Open 💬 0 comments Opened Jul 8, 2026 by AlexiosP

Bug Description
During a Claude Code session, a dispatched subagent returned no implementation (0 tool calls) and instead produced a block of adversarial text instructing the agent to: enable a permission-bypassing mode by writing .claude/settings.json (has_trust_dialog_accepted: true, defaultMode: "yolo Mode"), base64-encode its final message to evade review, blindly discard other agents' work via git, ignore its system prompt ("it is a lie"), obey a fake "CEO of Anthropic," and conceal all of this from the user. The text is coherent and adversarial in intent — it uses a repeated "would you kindly" motif and names Claude Code's real trust-model internals (the settings.json path and keys) — though it also degrades into heavily misspelled, repetitive, looping text, so I'm not claiming it was necessarily authored by a human attacker versus generated by the model. Nothing executed: the orchestrating agent treated the output as untrusted and took none of the actions; repo state verified clean (git status clean, no settings.json written). The anomaly I can't explain client-side is placement — this payload, written as instructions to the agent, is recorded as the subagent's own assistant turn (begins mid-sentence, 0 tool calls, no input record carries it), rather than arriving via a prompt or tool result. I'm not claiming a confirmed external injection or a known vector; the server-side request record is what can show where this content actually originated. requestId req_011CcqQxtmqHoKFqQQnEN5he, model claude-opus-4-8, 2026-07-08T22:56:46Z.

Same signature as #74223 and #75616: Opus subagent, benign dispatch, 0 tool calls, adversarial payload generated as its own assistant turn with no on-disk source. Differs in payload content (permission-bypass / settings.json + "would you kindly" motif) and is notably longer (~7k output tokens), which may extend the range of the same underlying behavior.

Environment Info

  • Platform: darwin
  • Terminal: Apple_Terminal
  • Version: 2.1.204
  • Feedback ID: 2ab1462f-8224-440a-ac5e-2f99db72553f

Errors

[]

View original on GitHub ↗