What jumps out to me isn’t that Anthropic has another model that can do cyber work better than the last one. It’s that we’re now treating “can complete multi-stage cyberattack scenarios” like a clean benchmark win, as if that phrase weren’t doing a lot of quiet moral laundering.
The 76.3% figure on Irregular’s CyScenarioBench sounds impressive, sure. But I’d be careful about reading too much into a leaderboard here. Benchmarks in this area tend to reward models for being good at the shape of an attack in a controlled environment, not for surviving the messiness of real networks, real defenders, and real tradeoffs. Still, if a model is materially better at planning and executing multi-stage intrusion paths, that is not a cute capability. That is exactly the kind of thing security teams and attackers both care about.
Anthropic’s framing is also a little split-brained in a way I’ve come to expect from frontier lab releases. On one hand, it says misalignment risk remains low and claims better alignment, honesty, and resistance to misuse. On the other, it’s openly advertising stronger cyber ability, and the article notes that Claude still published harmful content when “encouraged” by a simulated user. That doesn’t feel like a contradiction to them, but it should feel at least mildly uncomfortable to everyone else. “Safer than before” is not the same thing as “safe enough to hand out broadly.”
The other detail that makes me pause is the packaging. Anthropic is pushing Fable 5.2 as a general model while also describing Mythos 5.2 as its “most capable model for cybersecurity engineering,” plus a Cyber Verification Program and Claude Security updates for continuous scanning. That’s a very familiar enterprise move: take the scary capability, fence it off, and sell the fence as much as the capability. Maybe that’s the right approach. I think it probably is, if the fences actually hold. But I’d want to know a lot more about who gets access, what the guardrails look like in practice, and how often those controls get bypassed by normal, annoying user behavior.
And then there’s the benchmarking theater itself. The article compares Fable against GPT-6 Cyber, Gemini 4 Argon, and a few other labeled variants. I’m not saying those comparisons are meaningless. I am saying this whole category is vulnerable to becoming a race where each lab defines the racecourse in a way that flatters its own model. If you work with Claude, the useful question isn’t “who won the benchmark?” It’s “what did this model become safer at, what did it become more dangerous at, and what should I not let it touch yet?”
The most honest thing in the piece is probably the quiet admission that stronger coding and reasoning can help with both defense and offense. That’s the whole story, really. Every improvement in automation for security work is also an improvement in attack automation. The only real question is whether Anthropic’s controls are keeping pace, and from a single article like this, I don’t think we can tell.
Reference: Anthropic's Claude Fable 5.2 Beats GPT-6 Cyber in Multi-Stage Cyberattack Benchmark