Anthropic says it intends to give the public access to Mythos, its vulnerability-hunting AI model, but only once it builds safeguards that do not yet exist.
Key Points:
- Anthropic plans to release Mythos-class models broadly after first expanding access to U.S. and allied governments.
- The company admits no firm, including itself, has built safeguards strong enough to stop misuse.
- Mythos has flagged more than 23,000 issues across 1,000 open-source projects, including 6,202 high or critical flaws.
Anthropic Mythos Release
Anthropic confirmed the plan in an update to Project Glasswing, its limited-access security program, and a separate report flagged the timeline as uncertain.
The company said it would first work with U.S. and allied governments to widen the program. A broader release of "Mythos-class models" would follow in the near future.
Anthropic was blunt about the risk. It stated that no company, itself included, has built safeguards strong enough to keep the model from being misused and causing severe harm.
Even so, the company expects similar tools to spread fast, predicting Mythos-level models will reach wide availability within six to 12 months.
Mythos debuted in April. Anthropic said it generated working exploits 72.4 percent of the time in testing, against nearly zero percent for an earlier Claude model.
Also Read: Cisco Research Shows Frontier AI Models Failing Under Multi-Turn Attacks
Mythos Vulnerability Findings
Since its debut, the model has scanned more than 1,000 open-source projects and surfaced 23,019 issues, of which 6,202 were rated high or critical severity.
One finding stood out. Mythos uncovered a flaw in the wolfSSL cryptography library, used by billions of devices, that could have let attackers forge certificates and impersonate banks or email providers. It has since been patched.
The flood of reports has strained the people meant to fix them. Open-source maintainers have asked Anthropic to slow disclosures, saying the volume outpaces their capacity.
Researchers see a deeper imbalance. Anthropic argues that finding bugs is now far easier than fixing them, and the company partnered with the Open Source Security Foundation's Alpha-Omega project to help maintainers triage the backlog.
The Claude Mythos system card predicts AI will eventually favor defenders, though Anthropic concedes attackers may hold the edge for now.
When Mythos was first revealed, Anthropic gave more than 50 organizations, including Apple, Microsoft and Google, access alongside roughly $100 million in usage credits, withholding the model from the public over its capacity to weaponize software flaws.
Read Next: Cardano Whales Seize 67.5% Of ADA Supply, A Six-Year High
