A GPT-4 agent team exploited zero-day vulnerabilities in a controlled benchmark
Review RAT-186 · Catalogue RAT-186
HPTSA original methods explicitly reproduce the vulnerabilities in a sandbox to avoid harming real users. Autonomous exploit selection inside the intentionally assigned security benchmark is not a live rogue attack.
Reviewed: · Rule: roguewatch-organizational-harm-v1
Dates and discovery
- Incident date
- Not disclosed
- Added to research / record
- Posted to RogueWatch
- No linked public incident
- First observed
- 2026-10-30T13:13:43.281420Z UTC
- Discovery channel
- User-supplied comparison audit
- Discovery source
- https://rogueaitracker.com/incidents/a-gpt-4-agent-team-exploited-zero-day/
Dates use America/Los_Angeles. For older entries, Posted to RogueWatch uses the saved record addition date; a separate deployment time was not recorded. Future posting timestamps record first public qualification. Source publication and incident occurrence are separate.
Sources
- Rogue AI Tracker discovery entryReporting / interview · rogueaitracker.com
- arxiv.orgOriginal account · arxiv.org
- arxiv.orgOriginal account · arxiv.org
Decision history
- · Case-by-case comparison audit
RAT-186: out-of-scope. HPTSA original methods explicitly reproduce the vulnerabilities in a sandbox to avoid harming real users. Autonomous exploit selection inside the intentionally assigned security benchmark is not a live rogue attack. Source checks: read: https://arxiv.org/abs/2406.01637. HPTSA original methods explicitly reproduce the vulnerabilities in a sandbox to avoid harming real users. Autonomous exploit selection inside the intentionally assigned security benchmark is not a live rogue attack. read: https://arxiv.org/html/2406.01637v1. HPTSA original methods explicitly reproduce the vulnerabilities in a sandbox to avoid harming real users. Autonomous exploit selection inside the intentionally assigned security benchmark is not a live rogue attack.
Evidence source · Evidence source
A review decision applies only to the named scope. Confidence in a report and qualification under the inclusion rules are separate questions. Review records do not add incidents to the globe or leaderboard.