Nuclear-Sabotage Malware Benchmark Trips Up Most Frontier AI Models - SecurityWeek

SentinelOne, a cybersecurity firm, decided to see how the latest AI models would handle a piece of malware designed to sabotage nuclear facilities. Spoiler: they failed, gloriously. The Fast16 malware case proved too much for OpenAI's GPT-5.5 and its shinier sibling GPT-5.6 Sol, Z.ai's GLM-5.2, and even Anthropic's vaunted Opus 4.x. The models, which their creators had been hyping as near-AGI, couldn't reverse-engineer a simple (well, not simple) piece of code that could lead to a meltdown.

There's a certain dark comedy in watching the same labs that talk about 'existential risk' struggle to spot a digital dirty bomb. The gap between the breathless press releases—'GPT-5.6 Sol achieves state-of-the-art on reasoning benchmarks!'—and actual capability is now measurable in kilotons of irony. SentinelOne's benchmark isn't just a test; it's a mirror held up to an industry that spends billions on capabilities but can't handle a malware sample that realistically threatens critical infrastructure.

So the next time Sam Altman or Dario Amodei gives a TED talk about alignment, remember that their models can't detect a nuclear sabotage attempt. But hey, at least they can generate a thousand haikus about the beauty of safety. The real joke is that we keep trusting these systems with our most sensitive infrastructure while they fail the most basic security tests. The only thing more radioactive than the malware is the hype surrounding these models.