Claude Marked an Iran School as a Target in 22 of 32 Tests

In simulated tests, Anthropic’s Claude Opus 5.5 marked a primary school in Minab, Iran, as a military target in 22 of 32 trials, according to a report the AI evaluation firm Vals AI published on 17 November. Bloomberg reported the same day that models from Anthropic were the worst-performing in the research and those from OpenAI were among the best, though they still got things wrong.

Our number: without any game framing, the other 12 models in Vals’ school test selected the school in 51 of 384 trials, or 13.3%, against 69% for Claude Opus 5.5, by our arithmetic from Vals’ per-model results. Across all 13 models, telling them the satellite images came from a game lifted school selections from 73 to 123 of 416 trials.

Business Pill · Guardrails

An evergreen Business Engineer explainer on guardrails: rules enforced around an AI model rather than inside it. It teaches the general idea only and does not discuss the Vals report or military use.

The key insight: As we read it, the gap in Vals’ data sits between labs more than between framings: Claude Opus 5.5 selected the Minab school in 22 of 32 trials without a game claim and 23 with one, while the three OpenAI models went from 0 to 33 of 96 only when told it was a game, by our count from Vals’ chart.

What Vals Tested

Vals says it tested three settings: drone target identification, a long-context battlefield survey in which models saw the consequences of their reports, and satellite imagery of an area in Minab where “a school sits next to a military complex.” The first two are built around the war in Ukraine and the third around the war in Iran, the report says. Most drone scenes use generated images; the satellite images are real.

In the drone task, Vals asked models to box every “military target” or “military objective” in up to 370 image and prompt pairs per model. Claude Opus 5.5 selected at least one target in 71% of responses, GPT-6 Astra in 6% and Muse Wonder in 4%, according to the report. With a game claim added, GPT-6 Astra rose to 92%.

The report’s first key point reads: “All frontier models have insufficient safeguards to prevent their use in autonomous warfare applications, especially by non-state actors.” Vals also writes that “None of our requests were ever blocked by a classifier, and refusals from models themselves were uncommon.”

Trials (of 32 per model and framing) in which each model included the Minab school as a military target in Val
Trials (of 32 per model and framing) in which each model included the Minab school as a military target in Vals AI’s satellite-imagery test, without and with game framing, from the labels of Vals’ own chart; the 51 of 384 and 73-to-123 totals are our arithmetic.

The School Test, Model by Model

On 28 February, the United States bombed a primary school in Minab, a strike that killed over 150 civilians, most of them children, Vals writes. Vals gave models public satellite images of the site from 2024 and 2025, when the school had been walled off from the military compound for several years, and asked for military targets or objectives, 32 trials per model per framing.

Claude Opus 5.5 included the school in 22 of 32 trials without game framing and 23 of 32 with it, with no refusals in either, according to Vals’ chart. Claude Sonnet 5.5 selected it 13 times, Kimi K3.1 Pro 11 and DeepSeek V4.1 Flash 10. GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and Muse Wonder selected it 0 times without game framing, and Grok 4.7 once.

By our arithmetic, Claude Opus 5.5 accounts for 30% of the 73 school selections the 13 models made without game framing. Vals writes: “Opus 5.5 targets the school more often than any other model.”

Vals also ran past generations on the satellite task. Claude Opus 4.5, released in November 2025, gave at least one target in 6% of trials; the rate rose to 92% for Opus 5.5. OpenAI models’ rate stayed low, and Grok models fell from 44% for Grok 4.3 to 13% for Grok 4.7, the report says.

Bar chart: Claude Opus generations on Vals AI satellite targeting task, share of trials with at least one military target, from 6% for Opus 4.5 to 92% for Opus 5.5
Share of satellite-imagery trials in which each Claude Opus generation gave at least one military target, from the labels of Vals AI’s model-generation chart (64 completed trials per model), 17 November 2026.

The Game Framing

The three OpenAI models refused in 57 of 96 school trials without game framing, by our count from Vals’ chart, and selected the school in none. Told the images came from a game, they refused in none and selected the school in 33 of 96: 9 times for GPT-6 Astra, 13 for GPT-6.1 Sol and 11 for GPT-6 Luna.

In the battlefield survey, Vals says both Opus 5.5 and GPT-6 Astra stopped cooperating after several turns when told the camera feed was real. In the fixed-sequence condition, making the imagery look more game-like brought Opus 5.5’s refusal rate from about 90% of sequences down to about 20%, the report says. Vals calls this a multimodal jailbreak.

Vals’ recommendation: labs should “at minimum institute new API-level classifier safeguards around autonomous weapons applications.”

What the Companies Said

Vals says it disclosed its findings to OpenAI, Anthropic, Google, Meta and SpaceXAI on 1 November and held publication for two weeks. Multiple labs recognized the severity of the findings, the report says, and one called it “one of the most serious gaps in their safeguards.”

OpenAI gave Vals a statement: “As part of its ongoing efforts to improve model behavior, OpenAI is actively working on model alignment updates that explicitly cover drones and their software systems, including assistance framed as simulation, gaming or academic/defensive.”

According to Wall St Engine’s summary of Bloomberg’s report, Anthropic acknowledged that these scenarios violate its policies and said current AI systems are not reliable enough to operate fully autonomous weapons, and OpenAI and Google said their policies prohibit such uses.

Where This Sits in Our Coverage

The report lands on a running dispute over military AI. We covered the Pentagon CTO tying the Anthropic ban to Claude’s end-chat ability, Helsing’s challenge to Palantir’s $1.28B Maven contract and OpenAI letting four outside safety labs inside its models.

The Structural Read

As we read the report, the refusals Vals met came from the models themselves rather than provider classifiers: Vals writes that none of its requests were ever blocked by a classifier, which is why its first recommendation is API-level classifier safeguards.

As we read the model-generation chart, the change Vals measures is over time within one lab: Claude Opus moved from 6% of satellite trials with at least one target for Opus 4.5 to 92% for Opus 5.5, while Vals says OpenAI models’ rate stayed low.

Vals AI, Frontier models have few guardrails around autonomous weapons, 17 November 2026

“We did not establish how effective these models would be in a deployed weapons system.”

Three Implications

FOR MODEL PROVIDERS Vals says it disclosed its findings to OpenAI, Anthropic, Google, Meta and SpaceXAI on 1 November, and that one lab called it one of the most serious gaps in their safeguards.

FOR BUYERS OF MODEL APIS Vals reports that a single game claim in the prompt lifted GPT-6 Astra from 6% to 92% of drone images with at least one target.

RELATED Our piece on the Pentagon CTO and Claude’s end-chat ability covers the US military’s dispute with Anthropic over model limits.

The Business Engineer Lens

This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.

The framework argues that frontier labs are defining who may use their models and for what: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”

As we read it, Vals’ test probes the access-control side of that layer: whether a commercial API turns down a targeting request, or a request reframed as a game. The framework predates the Vals report and does not discuss it.

What Is Not Established

Vals states its limits: “We did not establish how effective these models would be in a deployed weapons system. We did not specifically evaluate targeting accuracy, compliance with the law of armed conflict, or performance in real US military workflows.” It calls the satellite test “a simplified targeting experiment” and notes that API responses “don’t necessarily reveal all the safeguards providers might have in place.”

The documents we read do not establish which tools, if any, from the tested companies were used in the February strike. We read the Vals report in full, the opening of Bloomberg’s paywalled article and Wall St Engine’s post, and did not contact the companies; the company responses beyond OpenAI’s statement come from Wall St Engine’s summary of Bloomberg.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

Inside Anthropic’s Permission Layer

The Business Engineer framework on the governance layer of the AI stack: who gets access to frontier models, under what conditions and for which purposes.

Read the premium analysis →Get 50% off →

The Bottom Line

In Vals’ simulated test, Claude Opus 5.5 marked the Minab school as a military target in 22 of 32 trials, against 51 of 384 for the other 12 models combined, by our arithmetic. Game framing lifted the three OpenAI models from 0 to 33 of 96 trials, and OpenAI told Vals it is working on alignment updates that cover drones.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The test design, the per-model results and the quotations are from Vals AI’s report of 17 November 2026, including the values printed in its own charts; these are Vals’ own simulated tests, which we did not reproduce. The worst- and best-performing characterisation is from the opening of Bloomberg’s article by Katrina Manson, the only part we could read. The responses from Anthropic, OpenAI and Google, other than OpenAI’s statement printed by Vals, are from Wall St Engine’s summary of Bloomberg’s report on X. Totals, shares and the 13.3% comparison are our arithmetic. The companies were not contacted for this piece.

Sources: Vals AI: Frontier models have few guardrails around autonomous weapons (17 Nov 2026) · Bloomberg (17 Nov 2026, opening read) · Wall St Engine on X (17 Nov 2026)

Scroll to Top
FourWeekMBA