Epoch AI argues Mythos’s cyber capabilities are more than hype, but uneven

Epoch AI argues Mythos’s cyber capabilities are more than hype, but uneven

Epoch AI published this assessment on June 11, 2026. It argues that Claude Mythos Preview made a large advance in exploit development: building a way to use a known software weakness to cause unauthorized behavior. Epoch AI says its Cyber-ECI places Mythos Preview 7 months ahead of the trend since early 2025, and that Anthropic’s analyses found it better at developing exploits that enable arbitrary code execution. The case for a comparable leap in finding vulnerabilities is less clear. Epoch AI points to a sharp rise in serious vulnerabilities found by Project Glasswing participants, but says the initiative’s spending may explain some of that increase. It also reports that Mythos Preview found just one low-severity vulnerability in curl, alongside four false positives. Epoch AI says earlier AI was already good at finding vulnerabilities, while Mythos Preview’s lower false positive rate and better prioritization may reduce the human effort needed to assess them. Epoch AI says Mythos 5 is modestly better still, though real-world usage reports remain to be seen. It concludes that if the capabilities were widely available, vulnerabilities would need to be patched much faster to prevent a big increase in successful cyberattacks.

Publication: 9ab3b0c9-7080-f5c4-77ed-39315e94d207

attribution — pass: The reading consistently attributes its claims to Epoch AI, the publisher of this single analytical post. It separately attributes the exploit-development findings to Anthropic’s analyses and the curl finding to the curl maintainers' context. It does not present source assertions as established fact.

So our current take on the public evidence is this: Mythos Preview was clearly a large improvement in exploit development

summary — uncertain: The overall shape is faithful: a strong exploit-development advance, an unclear vulnerability-discovery advance, and spending as a confounder. However, several central terms go unexplained and the definitions list is empty. Cyber-ECI, the key evidence, is never described as Epoch’s aggregate index built from about fifteen cyber benchmarks. Project Glasswing is not identified as Anthropic’s initiative giving partners access to Mythos Preview. Curl is not described as a heavily audited code library. Arbitrary code execution and false positives are also unexplained. Smallest repair: add brief plain-language glosses or tooltips for these terms.

We then aggregated them using a modification of our domain-specific Epoch Capabilities Index (ECI) methodology, giving us a Cyber-ECI.

date — pass: June 11, 2026 matches the UTC calendar day of evidence.originalPublished, which is stated as an instant from a feed backfill. Forecasts are reported conditionally, there is no hindsight framing, and the kind is publisher-page, so no archive disclosure is needed. The source says Mythos 5 was released on Tuesday; the reading does not repeat this.

support — uncertain: Most claims match their quoted text. However, the summary and claim 4 drop the source's hedge: the source says the capabilities would likely move us into a new regime. The reading instead says Epoch concludes vulnerabilities would need to be patched faster, which erases a stated uncertainty. Separately, the summary says Mythos 5 is modestly better still without the source's scope, which is exploit development. Smallest repair: restore the word likely and scope the Mythos 5 comparison.

If made widely available, these capabilities would likely move us into a new regime of cybersecurity

media — pass: No media was selected and no verified media candidates were supplied, so there is no identity or relevance problem.