GuidesHosting choices

Claude Code "safeguards flagged this message": what it means

Opus 5.5 refused with Details: [reasoning_extraction], or flagged a cyber or bio topic: what triggers each, how to find it, and the fallback model.

November 19, 2026The Everpod team
The short answer

A safety classifier on the model declined the request. The line under the message says which kind of flag it was. With Details: `[reasoning_extraction]`, the classifier read the request as asking the model to write out its internal reasoning, and the usual cause is an instruction somewhere in what Claude Code sends: a <thinking> or scratchpad section, a reasoning field in JSON output, a running log of the model’s thoughts. It can sit in CLAUDE.md, a skill, a subagent prompt, an output style or an MCP tool description as well as in your prompt. Run claude --safe-mode and send the same prompt: if it now goes through, a customization is the trigger. Reword or remove it, then start a new session, because the check reads the whole conversation and --continue or --resume carries the flagged content back in. A cybersecurity or biology flag is different: on Opus 5.5 Claude Code normally re-runs the request on another model by itself.

The messages, and which flag each one is

Claude Code’s error reference sorts these by wording, as of November 2026:

The full refusal users report from claude -p on Opus 5.5, in anthropics/claude-code#108048, reads:

API Error: Opus 5.5's safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with Opus 5.5.
Details: `[reasoning_extraction]`
Copied

The Details tag is the API’s refusal category. Anthropic’s refusals page names five: cyber (could enable cyber harm, and “Benign cybersecurity work can also trigger this category”), bio (could enable biological harm, and beneficial life sciences work can trigger it too), frontier_llm (could help develop competing AI models), reasoning_extraction and general_harms (another usage-policy area). The refusal is a normal response that stopped with stop_reason: "refusal", not an outage, so retrying the identical request does not change it.

Fixing a reasoning_extraction refusal

The trigger is wording, and Anthropic lists the shapes it usually takes: a <thinking>, <reasoning> or scratchpad section the model fills in before answering; a reasoning, thinking or trace field in JSON output or a tool input; the model’s private notes or a running log of its reasoning; reasoning asked for verbatim or in full. A pipeline that parses a reasoning field out of structured output is the classic case.

  1. Find where the wording lives. claude --safe-mode starts a session with CLAUDE.md, skills, MCP servers and hooks disabled. Send the same prompt there. If it goes through, search those files for instructions about thinking, reasoning, scratchpads or notes; if it is refused again, the prompt itself is the trigger.
  2. Reword it. Ask for a short explanation, the evidence behind a result, or a summary of the actions taken, which the docs say is still fine. To read the model’s thinking yourself, Claude Code’s showThinkingSummaries setting shows summaries of it.
  3. Start a new session. A customization is read when a session starts, and the flagged turn stays in the transcript. In an interactive session, Esc twice or /rewind steps back to before the turn; /clear starts over in the same project.

The refusals page says this category “has no recommended fallback model, so change the prompt rather than retry the request. If a reworded request is still declined, contact support and include the request ID.” The refusal message carries a Request ID and a Message ID for that.

When nothing you wrote asks for reasoning

The open reports describe this case. In #108048 the refusal came right after the agent read a long notes file, the reporter narrowed it to a 30-line block of that file, and it happened only when user-scope rules files were loaded. In #108720, headless coding agents on Opus 5.5 were refused in two of three runs, each time after reading different source files. These are users’ reports with no maintainer reply as of November 20, 2026, but they fit what the docs say: the check reads everything in the conversation, file contents and loaded context included, and the first request of a session already carries your CLAUDE.md and git status. So the trigger can be text the agent read, not text you typed. If the same task keeps tripping on the same file, that file is worth reading for anything shaped like the list above. Where you believe the refusal is wrong, /feedback or support with the request ID is the route Anthropic gives.

Cybersecurity and biology flags, and the automatic switch

Opus 5.5, Sonnet 5.5, Opus 5 and the Fable models run classifiers that, per the model configuration docs, most often flag cybersecurity and biology content. From Claude Code 2.1.219, a flagged request is re-run on a fallback model: on Opus 5.5, biology-flagged requests re-run on Opus 5 and cybersecurity-flagged requests on Opus 4.8. The session then stays on the fallback model until you pick another with /model. Security research and biology codebases trip this often, sometimes on the first request; Anthropic calls that “expected routing for these domains, not an account flag.”

The switch does not happen, and you see the refusal instead, when:

For legitimate security work that keeps being flagged, the route Anthropic offers is its Cyber Verification Program, linked from the cybersecurity message; for organizations, the docs point to the Anthropic account team about trusted access programs.

Headless runs and background subagents

A refusal costs more where nobody is watching. In claude -p there is no rewind, and the docs’ advice is to retry “with a rephrased prompt in a new session without --continue”. A job that resumes the same session on failure will meet the same refusal every time, so treat a refusal as a stop that needs a changed prompt, not as a transient error to retry; the guide to running Claude Code headless on a schedule covers checking the exit and output of such jobs.

A subagent never pauses to ask. With automatic switching on (the default), its flagged request re-runs on the fallback model; with Ask each time, or when there is no fallback, as with reasoning_extraction, the subagent fails with “Agent terminated early due to an API error” and the main session decides what to do about the unfinished task. A user in #105359 lost a two-hour background agent this way. If you write subagent briefs or skills that ask the agent to log or show its reasoning, that wording is the first thing to change.

Run Claude Code on an always-on developer pod.

A developer pod is a permanent cloud computer (VPS), set up and secured for you, with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed. Run more agents at once than your laptop can hold, and keep them working when it’s closed. From $24 a month, built in about ten minutes.