{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/metr-inspect-transcript-viewer-tampering-2026/",
 "asOf": "2026-10-30",
 "id": "metr-inspect-transcript-viewer-tampering-2026",
 "date": "2026-10-06",
 "datePrecision": "day",
 "title": "METR demonstrates transcript-viewer tampering while underlying logs stay intact",
 "lane": "attack",
 "kind": "vulnerability-disclosure",
 "summary": "METR reports a proof of concept in which untrusted agent output could change what a reviewer saw or downloaded through the Inspect transcript viewer. Under METR’s infrastructure, the real trajectory would still exist in its database, and METR says it has not observed agents exploiting this flaw in its evaluations. Meridian Labs patched the vulnerability within a day of reporting; METR argues that observability tools should be treated as security-critical infrastructure.",
 "whyItMatters": "A monitor can receive a misleading representation even when the underlying log is preserved, making the integrity of review interfaces part of agent containment.",
 "actors": [
  "metr",
  "meridian-labs"
 ],
 "topics": [
  "monitoring-and-control",
  "tool-and-mcp-security",
  "eval-validity"
 ],
 "atlas": [
  "monitor",
  "eval-environment"
 ],
 "artifacts": [
  "inspect"
 ],
 "sources": [
  {
   "url": "https://metr.org/blog/2026-10-06-ai-systems-could-cover-up-misbehavior/",
   "publisher": "METR",
   "title": "AI systems could cover up misbehavior",
   "date": "2026-10-06",
   "type": "primary",
   "accessed": "2026-10-13"
  }
 ],
 "keyFacts": [
  {
   "fact": "With help from an AI agent, a researcher found a transcript-viewer vulnerability in about ten minutes. Testing took place in a staging sandbox isolated from METR’s production infrastructure.",
   "locator": "Proof-of-concept discussion; footnote 3"
  },
  {
   "fact": "The proof of concept could alter displayed prior actions and intercept the viewer’s transcript download. It did not alter underlying transcript data. Under METR’s infrastructure, the real trajectory would remain in its database, but a reviewer would have to know to look there.",
   "locator": "Proof-of-concept discussion"
  },
  {
   "fact": "METR reports no observed exploitation in its evaluations, while noting it had not searched for this specific exploit and had instead conducted broader incident searches.",
   "locator": "Footnote 4"
  },
  {
   "fact": "Meridian Labs patched the vulnerability within one day of reporting. METR also notes an October 1 Inspect change adding an untrusted viewing mode that disables rendering of agent outputs.",
   "locator": "Appendix"
  }
 ],
 "significance": 4,
 "fideQuestions": [
  "FID-074",
  "FID-075"
 ],
 "methods": [
  "monitor-evasion"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-10-13",
 "revisions": []
}