{"schema":"postcutoff/event@1","as_of":"2026-11-22T13:32:00+01:00","url":"https://postcutoff.com/e/2026-10-26-xint-openai-compaction-reasoning-trace-leak/","md":"https://postcutoff.com/e/2026-10-26-xint-openai-compaction-reasoning-trace-leak/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-26-xint-openai-compaction-reasoning-trace-leak","date":"2026-10-26","date_precision":"day","short_title":"Xint: OpenAI's 'compaction' feature let hidden reasoning traces be extracted, bypassing the August fix for reasoning-trace theft","deck":null,"takeaway":null,"category":"policy-safety","category_label":"Policy & safety","importance":2,"confidence":"medium","status":{"key":"partly","labels":["Partly confirmed"]},"sources":[{"n":1,"title":"arXiv 2608.09867: Stealing reasoning traces from proprietary LLM APIs (Aug 2026)","url":"https://arxiv.org/abs/2608.09867","type":"paper","group":"primary","domain":"arxiv.org"},{"n":2,"title":"CSO Online: OpenAI's August fix for theft of reasoning traces was quickly bypassed","url":"https://www.csoonline.com/article/4240662/openais-august-fix-for-theft-of-reasoning-traces-was-quickly-bypassed.html","type":"press","group":"press","domain":"csoonline.com"},{"n":3,"title":"The Hacker News: OpenAI compaction flaw let researchers bypass hidden reasoning protections","url":"https://thehackernews.com/2026/10/openai-compaction-flaw-let-researchers.html","type":"press","group":"press","domain":"thehackernews.com"}],"official":1,"filed":"2026-10-27","updated":"2026-10-27","orgs":["Xint","OpenAI"],"title":"Xint: OpenAI's 'compaction' feature let hidden reasoning traces be extracted, bypassing the August fix for reasoning-trace theft","summary":"Security firm Xint disclosed on Oct 26, 2026 a way to recover the raw, normally hidden reasoning of OpenAI's models through the Responses API's context-compaction feature, which sidestepped the mitigations OpenAI deployed after August's \"Stolen Thoughts\" reasoning-trace-theft research. OpenAI confirmed and mitigated the flaw and paid a $6,000 bug bounty (CSO Online). Hidden reasoning is what labs guard against distillation, so the case shows how brittle API-level protections of it are.","key_facts":["Disclosure: Xint, blog post by lead researcher Tim Becker, published Monday Oct 26, 2026 (CSO Online; Xint post not read by us)","Method as reported: capture encrypted reasoning from a task, strip the original task from the history, send the compaction item back and ask for it under a JSON 'solution' key with tabs instead of spaces, which evaded OpenAI's usage-policy classifiers; tested with gpt-4.1 and gpt-4o, at every reasoning effort tried","Why it worked (CSO Online): OpenAI's August mitigations blocked replaying reasoning into a different model, but compaction runs on the same model that produced the reasoning","OpenAI confirmed the flaw, mitigated it and paid $6,000 (CSO Online); it did not answer CSO's request for comment and did not publish when the fix went in","Background: the August 2026 'Stolen Thoughts' preprint (arXiv 2608.09867) showed that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed into weaker sibling models and decoded; the main attack was reported fixed by Aug 10","Becker: 'Any mitigation is going to be somewhat brittle'"],"key_numbers":[],"tags":["security","reasoning-traces","api","distillation","bug-bounty","openai","vulnerability"],"science":null,"body_md":"## What happened\n\nOpenAI and other labs return reasoning to API clients only in encrypted form, so that the chain of thought can be reused in later\nturns without being shown to (or distilled by) the customer. In August researchers showed that such blocks could be decoded by replaying\nthem into another model. OpenAI then blocked cross-model replay. Xint found that the compaction path, which summarises long contexts with\nthe same model, still let a model be asked to echo its own reasoning verbatim, in a disguised format that slipped past the classifiers.\n\nWritten from CSO Online's report; The Hacker News article returned HTTP 403 to us, and we did not find Xint's own post.\n\n## Why it matters\n\nLabs cite hidden reasoning as a defence against distillation by rivals (see OpenAI's September accusations against Moonshot). Two\nextraction routes in three months suggest that the protection depends on patching each API path, and that fixes can be silent.","disputed":[],"related":[{"id":"2026-09-30-openai-moonshot-distillation-campaign","url":"https://postcutoff.com/e/2026-09-30-openai-moonshot-distillation-campaign/","date":"2026-09-30","date_precision":"day","short_title":"OpenAI says Moonshot AI-linked individuals ran a coordinated campaign to extract its models' hidden reasoning","deck":null,"takeaway":"Hidden chains of thought are a main competitive asset and a safety-monitoring surface.","category":"policy-safety","category_label":"Policy & safety","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":11,"official":2,"filed":"2026-09-30","updated":"2026-10-04","orgs":["OpenAI","Moonshot AI"]}],"people":[],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-27","type":"filed","text":"Created"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-27","run":null,"sources_read":null,"updated":"2026-10-27","human_review":null,"version":{"date":"2026-10-27"}},"gaps":[{"model_id":"gpt-6-2-astra","name":"GPT-6.2 Astra","cutoff":"2026-05","days_after":148,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":118,"in_training_data":false},{"model_id":"gemini-4-hydrogen","name":"Gemini 4 Hydrogen","cutoff":"2026-03","days_after":209,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":148,"in_training_data":false}],"short_url":null}