Safety

White House Demands Anthropic Fix Fable 5 Jailbreak Vulnerabilities

The White House has told Anthropic it must address jailbreak vulnerabilities in its Fable 5 model, citing NSA findings that guardrails can be bypassed.

Professional woman working on laptop in a server room, showcasing technology and remote work.

Photo: Christina Morillo / Pexels

The White House is pressing Anthropic to resolve security concerns surrounding its Fable 5 AI model, which was temporarily blocked due to potential jailbreak vulnerabilities. Federal officials have informed Anthropic that if it intends to re-release the model, it must take concrete steps to address the government’s concerns. The administration argues that the National Security Agency has confirmed methods to bypass the model’s guardrails, which are designed to prevent access to sensitive capabilities related to cybersecurity, chemistry, and biology. Anthropic has previously dismissed these concerns as overblown, asserting that the impact of jailbreaks is minimal. The company reiterated this stance during a technical meeting with the Commerce Department and the Office of the National Cyber Director, Sean Cairncross, on Monday. However, officials now insist that the focus has shifted from debate to action, with the administration viewing the issue as Anthropic’s responsibility to resolve.

Officials emphasize that neither the Commerce Department’s Center for AI Standards and Innovation nor the National Security Agency has the resources to investigate every possible jailbreak across all models. As a result, the administration is urging Anthropic to proactively test its frontier models, including Fable 5, to identify and report potential vulnerabilities. The White House believes that continuous testing is essential to ensure the models remain secure. However, independent cybersecurity experts suggest that guardrails are only a temporary measure, as skilled users and future AI models are likely to find ways to circumvent constraints. This raises fundamental questions about the feasibility of preventing jailbreaks entirely.

The White House spokesperson declined to comment on the matter. Meanwhile, the situation involving the Acting Director of National Intelligence, Bill Pulte, remains unresolved, with Trump delaying the confirmation of his permanent nominee, Jay Clayton, to prevent Pulte from taking office. Senate Republicans have indicated they will proceed with Clayton’s hearing unless the nomination is withdrawn. The controversy has highlighted concerns about Pulte’s lack of experience and his perceived eagerness for the role’s perks without fulfilling its responsibilities. Source: wired

Key points

The White House has told Anthropic it must address jailbreak vulnerabilities in its Fable 5 model.
The National Security Agency has confirmed methods to bypass the model’s guardrails.
Anthropic has dismissed the concerns as overblown, asserting that the impact of jailbreaks is minimal.
The administration is urging Anthropic to proactively test its frontier models to identify and report potential vulnerabilities.
Independent cybersecurity experts suggest that guardrails are only a temporary measure as skilled users and future AI models may find ways to circumvent constraints.

Source: WIRED Read the original →

WRITTEN BY

Nadia Rahman

AI Safety, Alignment & Policy

Nadia follows AI safety, alignment, regulation, and the policy debates shaping the field.