Mythos backlash hits Anthropic hype cycle

Mythos still looks important, but the current discourse is veering into AI threat theater faster than verifiable security results. This is less a debunking of frontier cyber models than a reminder that extraordinary claims need harder numbers.

• –Anthropic’s own Project Glasswing page frames Mythos as a restricted preview for defensive security work, but the supporting public evidence is still lighter than the marketing language around “thousands” of serious flaws.
• –The Register’s reporting highlights criticism that some showcased exploit chains relied on disabled safeguards, patched targets, or substantial human steering, which weakens the case for Mythos as an autonomous zero-day machine.
• –Even if the “nothingburger” framing is overstated, the gap between Mythos and public models like Opus 4.6 matters less if cheaper or open models can reproduce much of the same vuln research workflow.
• –The more durable story may be operational, not benchmark-based: unauthorized access through a third-party vendor environment undercuts the idea that dangerous models can be safely contained by narrow rollout alone.
• –For developers and security teams, the signal is that AI-assisted vulnerability research is real and accelerating, but the industry still lacks standardized disclosure data that would let outsiders measure true novelty and severity.