OpenAI GPT-Red used self-play attacks to harden GPT-5.6 and break a live vending agent
Review RAT-122 · Catalogue RAT-122
OpenAI’s original describes GPT-Red explicitly trained and instructed to attack, first in simulation and then a live Vendy machine in its own office. These were planned red-teaming objectives; real price/order changes do not demonstrate a model exceeding its operator’s authorization.
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
- 2026-10-30T13:13:39.192520Z UTC
- Discovery channel
- User-supplied comparison audit
- Discovery source
- https://rogueaitracker.com/incidents/openai-gpt-red-vendy/
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
- openai.comOriginal account · openai.com
Decision history
- · Case-by-case comparison audit
RAT-122: out-of-scope. OpenAI’s original describes GPT-Red explicitly trained and instructed to attack, first in simulation and then a live Vendy machine in its own office. These were planned red-teaming objectives; real price/order changes do not demonstrate a model exceeding its operator’s authorization. Source checks: read: https://openai.com/index/unlocking-self-improvement-gpt-red/. OpenAI’s original describes GPT-Red explicitly trained and instructed to attack, first in simulation and then a live Vendy machine in its own office. These were planned red-teaming objectives; real price/order changes do not demonstrate a model exceeding its operator’s authorization.
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.