{
  "id": 1272957,
  "title": "When \"Hidden\" Reasoning Isn't Hidden: The OpenAI/Anthropic/Google Encrypted Trace Replay Bug",
  "url": "https://urgent.news/2026/08/16/when-hidden-reasoning-isnt-hidden-the-openai-anthropic-google",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-16T13:04:27.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/coridev/when-hidden-reasoning-isnt-hidden-the-openaianthropicgoogle-encrypted-trace-replay-bug-2a5j"
  },
  "original_language": "en",
  "account": "In August 2026, researchers discovered a vulnerability in encrypted reasoning across OpenAI, Anthropic, and Google APIs. This flaw allowed encrypted reasoning objects to be replayed outside their original context, even across different users and into weaker models. The result was hundreds of secrets, such as API keys, passwords, and access tokens, being pulled out of reasoning traces that never appeared in the visible model output. These secrets were encrypted and designed to remain opaque, but the bug allowed them to be readable by unauthorized parties. The issue lies in the fact that the encryption lacks strict binding to session, user, and model identity, making it merely obfuscation rather than true security. Agentic workflows, which pull data into reasoning contexts that don't appear in the final answer, can further exacerbate the problem. Traditional LLM security tools focus on monitoring the visible input/output stream, overlooking the encrypted reasoning traces that sit outside this scope. Additionally, secret-scanning typically looks for secrets in text that reaches humans or logs, not in the encrypted reasoning traces themselves. The bug specifically enables the leakage of sensitive information through side channels, where weaker models decode and echo back the content of stronger models' reasoning traces. To address this, a dedicated secret and credential detection layer, Layer 4, would be necessary to identify and flag any leaked secrets embedded in the reasoning traces, regardless of the surrounding context.",
  "summary": "Researchers disclosed in August 2026 that encrypted reasoning objects across OpenAI, Anthropic, and Google APIs could be replayed outside their original context. Not just replayed by the same user in a later session. Replayed across users. Replayed into weaker models that had no business decoding a stronger model's internal reasoning at all. The result: hundreds of real secrets pulled out of…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}