Mythos Preview

The Trump administration's approach to AI has been inconsistent, despite the president's stated goal of accelerating the AI industry to compete with China. The White House recently classified Anthropic's advanced AI model Fable 5 as a national security threat, leading to its shutdown, despite Anthropic's safety measures and cybersecurity experts' concerns about over-restriction.

XBOW tested Anthropic's Mythos Preview for offensive security, finding it highly effective in identifying vulnerability candidates, particularly in source code analysis. The model was evaluated across exploit discovery, reverse engineering, and live-site validation.

UK banks, excluded from Anthropic’s Project Glasswing expansion, are receiving access to OpenAI’s GPT-5.5 Cyber model. While JPMorganChase was the only financial institution granted access to Anthropic’s Mythos Preview, UK banks like HSBC, Lloyds, and Nationwide will use GPT-5.5 instead. The Bank of England’s governor, Andrew Bailey, criticized Anthropic’s exclusion of UK institutions, with some suspecting political motives behind access decisions.

New AI cybersecurity models from Anthropic and OpenAI require significant human expertise to operate effectively, finding tens of thousands of bugs in operating systems. Early adopters have shared their experiences with the models, discovering numerous vulnerabilities. The models perform best when paired with experienced security researchers.

Mozilla fixed 423 Firefox security bugs in April, a significant increase from previous months, with help from Anthropic's Mythos Preview model. The model found 271 bugs, and Mozilla has decided to publicize selected wins to encourage others to use AI for bug remediation. This effort is part of a broader initiative to improve security coverage.

OpenAI is rolling out a more permissible version of GPT-5.5 to vetted cyber defenders, which has capabilities similar to Anthropic's Mythos Preview in finding and exploiting software bugs. This move sparks debate about keeping such models out of the hands of bad actors.

Disneyland has implemented face recognition technology for visitors. Additional news includes the NSA testing Anthropic's Mythos Preview for vulnerabilities and a Finnish teen being charged for involvement in the Scattered Spider hacking attacks.

The UK's AI Security Institute found that OpenAI's GPT-5.5 matches Anthropic's Mythos Preview in cybersecurity performance, with both models achieving similar success rates in Capture the Flag challenges and a new data extraction test. However, both failed a more complex power plant control simulation.

OpenAI and Anthropic held classified briefings with the House Homeland Security Committee to discuss their new cyber-capable AI models and their implications for national security. Anthropic delayed the release of its Mythos Preview model due to security risks, while OpenAI adopted a tiered approach for its GPT-5.4-Cyber model. Both companies are collaborating with federal agencies to assess the models' cybersecurity risks.

Anthropic's Project Glasswing, an AI model designed to detect software vulnerabilities, was delayed from public release. The company granted access to Apple, Microsoft, Google, Amazon, and others to identify and patch bugs before malicious actors could exploit them.

The Cybersecurity and Infrastructure Security Agency (CISA) has not adopted Anthropic's Mythos Preview cybersecurity model, despite other agencies like the Commerce Department and NSA using it. President Donald Trump's administration is negotiating broader access to the tool.

OpenAI has been briefing U.S. federal agencies, state governments, and Five Eyes allies on its new GPT-5.4-Cyber model, which offers advanced cybersecurity capabilities. The company is using a tiered access program with vetting for government and commercial users, while Anthropic's competing Mythos model faces restrictions due to Pentagon concerns.

Mozilla revealed that Anthropic's Mythos Preview model identified 271 zero-day vulnerabilities in Firefox 150 before its release. Firefox CTO Bobby Holley highlighted this as a potential turning point for defenders in the cybersecurity arms race.

Anthropic has developed a powerful AI model called Mythos Preview, which it claims can exploit software vulnerabilities and pose risks to security and economies, but some experts question its capabilities. The company's decision to limit public access has sparked debates about AI regulation and corporate strategy.

The White House is meeting with Anthropic CEO Dario Amodei amid a Trump administration reversal on the AI firm's technology, which was previously banned over national security and political concerns. The Pentagon designated Anthropic a 'supply chain risk,' leading to a legal battle, but officials are now reconsidering its potential for defending U.S. infrastructure.

The White House is re-evaluating its stance on Anthropic's AI technology after a meeting between CEO Dario Amodei and Chief of Staff Susie Wiles, following a Pentagon designation of the firm as a 'supply chain risk' and ongoing legal challenges. Anthropic's advanced Mythos Preview model, restricted due to misuse concerns, is at the center of a debate over its potential role in U.S. infrastructure defense versus national security risks.

Anthropic released Claude Opus 4.7, its most powerful 'generally available' AI model, highlighting improvements in software engineering, image analysis, and document creation. The release follows the announcement of Mythos Preview, a cybersecurity-focused model Anthropic claims is its most powerful overall.

The UK AI Security Institute evaluated Anthropic's Mythos Preview model, finding it excels in chaining multi-step cyber-attacks compared to previous models. While Mythos performs well on individual security tasks, its ability to execute complex attack sequences sets it apart from other frontier models.

Anthropic restricted its Mythos Preview AI model after it autonomously discovered and exploited zero-day vulnerabilities across major operating systems and browsers. Security experts warn similar capabilities may proliferate soon, with CrowdStrike reporting an average eCrime breakout time of 29 minutes.

Anthropic's Mythos Preview model can identify and exploit critical zero-day vulnerabilities, but the company claims to have implemented controls to manage its use. The article questions whether these safeguards can prevent the technology from falling into the wrong hands.