Anthropic widens access to Claude Mythos AI model for cyber defense
In short
Anthropic said it will broaden who can use Claude Mythos, its most advanced AI model, which had been limited to selected organizations because of the risk it could be misused in cyberattacks. Use is restricted to defensive cybersecurity purposes, and the company is merging two access programs into a single structure with three levels of permission. Nikkei reported that vetted individual experts and small and medium-sized businesses will also be able to apply.
Read the full story 1 min read
Anthropic announced on October 6 US time that it will expand who can use Claude Mythos, its most advanced AI model, which the company had kept to a limited set of organizations because of the risk it could be misused in cyberattacks. The company said use is limited to cybersecurity defense purposes. [ 1 , 2 , 3 ]
Nikkei reported that Anthropic will allow vetted individual experts and small and medium-sized businesses to use the latest model, restricted to cyber defense. The model is the company's top-tier product and has high performance in the cyber field, which is why it had been offered on a limited basis, the newspaper said. It added that the corporate consortium able to use Mythos, Project Glasswing, is being broadened into the wider Cyber Verification Program. [ 1 ]
Impress Watch reported that the company is merging Project Glasswing and the Cyber Verification Program and setting three access levels. Defense Access covers incident response, malware reverse engineering and vulnerability analysis. Red Team Access adds approved penetration testing and red-team activity, while operations that could cause large-scale disruption, such as ransomware deployment or damage to physical systems, remain blocked. The least restricted tier, Specialized Access, is for a limited set of organizations authorized to test systems including aviation operations, power grids, telecommunications networks, interbank transfer infrastructure and government administrative networks. Existing Project Glasswing participants move to Specialized Access without re-review. [ 3 ]
According to Impress Watch, the program is available through Claude Platform, Google Cloud Vertex AI and Microsoft Foundry, and on Amazon Bedrock only for users covered by Enterprise Frontier Safeguards. Participating organizations are required to retain data so misuse can be monitored, and when Enterprise Frontier Safeguards launches later in the autumn, they will be able to store that data in cloud infrastructure they manage themselves. [ 3 ]
Why it matters
The documents present Claude Mythos as a top-tier model with strong cyber capabilities that Anthropic had deliberately kept to a small group of organizations because of misuse risk. Extending it to more security specialists, and according to one report to individuals and smaller firms, changes the pool of users who can deploy high-capability tooling for defense and approved testing. Impress Watch says safeguards remain in the form of tiered access levels, blocked operations and data retention for misuse monitoring.
Key facts
- Anthropic announced on October 6 US time that it will expand availability of its Claude Mythos AI model. [ 1 , 2 , 3 ]
- Use of the model is limited to cyber defense purposes and to vetted or approved organizations. [ 1 , 2 , 3 ]
- The model had been restricted to a limited set of companies because of the risk it could be abused in cyberattacks. [ 1 , 2 ]
- Nikkei reported that vetted individual experts and small and medium-sized businesses will be allowed to use the latest model for cyber defense. [ 1 ]
- Impress Watch reported the company is merging Project Glasswing and the Cyber Verification Program and creating three access levels: Defense Access, Red Team Access and Specialized Access. [ 3 ]
- The most restricted-redaction tier, Specialized Access, covers a limited set of organizations authorized to test systems including aviation operations, power grids, telecom networks, interbank transfer infrastructure and government administrative networks. [ 3 ]
- CVP is available through Claude Platform, Google Cloud Vertex AI and Microsoft Foundry, and on Amazon Bedrock only for Enterprise Frontier Safeguards users. [ 3 ]
- Participating organizations are required to retain data so misuse can be monitored. [ 3 ]
Confirmed by several sources
- Anthropic announced it is expanding availability of its Claude Mythos model, previously limited to selected organizations. [ 1 , 2 , 3 ]
- The expanded availability is tied to cybersecurity defense purposes. [ 1 , 2 , 3 ]
- The restriction had been in place because the model could be misused in cyberattacks. [ 1 , 2 ]
- Anthropic disclosed the move on October 6 US time. [ 1 , 3 ]
Still unclear
- Whether individual experts and small and medium-sized businesses are eligible for access. Only Nikkei states this; NHK and Impress Watch describe expansion to security specialists and organizations without mentioning individuals or smaller firms.
- The exact announcement time in Japanese local time. The documents say only October 6 US time, with no clock time.
- How many organizations will be admitted under each access level. The documents give no participant numbers.
- The full scope of the conditions NHK describes for permitted use. NHK's text is truncated in the available document.