AI model
Preview-release AI model from Anthropic, tested on agentic and security tasks such as finding software vulnerabilities and chaining them into working exploits.
Claude agents designed 1,440 protein binders and 354 bound. The interesting part is that the prompts, provenance and every measurement went out with the number.
Reality
Cloudflare pointed Anthropic's Mythos Preview at more than fifty of its own repositories and watched it write, compile and run its own proofs of exploitability. Its refusals on identical code did not repeat.
Reality
A researcher ran the same FreeBSD scan through base and abliterated open-weight builds and found the uncensored ones graduating three to four times as many findings, while the most aggressive one never surfaced the actual CVE.
Reality
The hosted Claude Security beta bills as ordinary Claude usage with no platform fee, while Mythos 5 stays with vetted partners. What runs the Enterprise scans is not settled by the sources.
Reality
Its own Risk Report says an internal flag that disabled blocking also disabled logging, on a surface staffed by vendors that could not screen out CB-1 threat actors.
Reality
The vendor revised the rating upward and cited its own cybersecurity incidents rather than theory. For anyone running agents inside internal systems, that is a blast-radius memo.
Reality
Anthropic's own red team reports identical agents sabotaging each other on a shared job, and colluding on price floors in a separate game. Single-agent evals will not catch either.
Reality