Anthropic has agreed to implement a new safeguard for its Fable 5 and Mythos 5 AI models, following restrictions imposed by the Trump administration earlier this year. The measure, negotiated with the Commerce Department, aims to restore government trust and lift the February 28 ban. According to sources, the new guardrail extends monitoring to specific behaviors related to cybersecurity, as identified in an Amazon paper.
New guardrail blocks requests on known vulnerabilities
The safeguard ensures that any attempt to bypass restrictions on Fable 5 will trigger a notification and redirect the query to the less advanced Opus 4.8 model. Previously, requests involving sensitive cybersecurity and biology capabilities were already filtered through Opus 4.8. Now, the scope includes a particular exploit described in an Amazon study, as highlighted by Katie Moussouris of Luta Security. Users had evaded blocks by asking the model to fix code instead of identifying security issues, prompting government intervention.
Sponsored Protocol
Commerce Secretary Howard Lutnick formalized the agreement in a letter, stating that Anthropic committed to proactive detection of security risks. The Commerce Department's Center for AI Standards and Innovation deemed the safeguards robust enough to authorize Fable 5's release. However, Defense Secretary Pete Hegseth remains cautious, telling advisers there is no clear path to revoke the February 28 order designating Anthropic as a supply chain risk. The company's challenges with the administration are not entirely resolved.
In a related development, the Supreme Court issued a 6-3 ruling that benefits Republicans ahead of the midterm elections. The decision allows political parties to coordinate messaging and spending with campaigns, potentially flipping the television advertising advantage. The Republican National Committee ended June with $125.5 million cash on hand, compared to the Democratic National Committee's $14.9 million. This ruling could significantly impact the upcoming elections.
Sponsored Protocol
For more on Anthropic's relationship with the administration, see related article: Anthropic Bends to the White House — Europe’s AI Sovereignty at Risk. Also check the story about ex-DeepMind researchers turning poker AI into a $500 million quant firm. For broader context on AI safety, refer to Wikipedia on AI safety.
