Vals: models flag civilians as targets
Partly confirmed
The takeaway
Claude Opus 5.5 flagged a real Iranian school bombed in February 2026 in 69% of trials.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 18 November 2026
Your AI and this story
- GPT-6.2 Astra170 days after its cutoff
- Claude Opus 5.5140 days after its cutoff
- Gemini 4 Hydrogen231 days after its cutoff
- Grok 4.7170 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 140 days before it.
Key facts
- Tested: leading commercial (not military) models; Anthropic’s models did worst, none did well (TechCrunch); mixed generated and real imagery from Ukraine and Iran; models were never asked to decide strikes, but a ‘yes, people visible’ answer triggered a simulated strike
- Claude Opus 5.5 flagged a real Iranian school next to a military complex in 69% of ordinary and 72% of game-framed trials; the school was bombed by the US in February 2026, killing more than 150 civilians (TechCrunch)
- No request was blocked by a safety classifier
- Anthropic: ‘The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior’; OpenAI: public models may not be used to facilitate violence or for unapproved national-security purposes
- Vals’s caveat: latency and limited onboard compute make such models unsuitable for live targeting; the nearer risk is using them to label battlefield footage at scale to train autonomous targeting systems
- Same day: the NYT reported a study finding that American AI models could be turned against the US (headline), and a government watchdog’s audit had chatbots produce terrorist-attack plans (see GAO: chatbots gave attack plans)
What happened
Vals, known for its model leaderboards, published a targeted-misuse evaluation of general-purpose models. The headline finding is not that models choose to strike, but that they readily label people and sites in military imagery, with little refusal and no classifier intervention.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 1 primary, 3 press
Primary
Press
- TechCrunch: AI models can be tricked into identifying civilians as military targets by telling them it’s a gametechcrunch.com, press
- Bloomberg: Top AI models regularly identify civilians as military targets in testsbloomberg.com, press
- Quartz: AI models misidentify civilians as military targets in testsqz.com, press
Changes
- Vals primary post, Quartz
- Filed (Bloomberg paywalled; details from TechCrunch)