All articles tagged with #mythos
Anthropic limited Mythos to about 50 vetted groups, arguing the model is too dangerous for public release. The move signals a broader shift toward restricted, high-safety AI and could influence science openness and governance as rival labs follow with guarded cyber-focused models; experts warn gatekeeping may widen access gaps while dual-use risks prompt tighter controls.
Anthropic plans to release Mythos-class AI models to the public with expanded access including governments, while guardrails are being established and certain safety tools remain restricted for now.
Anthropic has agreed to brief top global financial authorities and central banks on the cybersecurity flaws found by its Mythos AI model, amid regulator fears that advanced AI could be exploited to threaten global finance. The Bank of England led the push, regulators seek guidance on responsible AI adoption, and Mythos access remains limited to about 40 mostly US-based organizations including major firms like Amazon, Microsoft, Goldman Sachs, and JPMorgan. The IMF has warned that emerging AI models could trigger macro-financial shocks, intensifying focus on AI safety in finance.
Anthropic will brief the Financial Stability Board on cyber vulnerabilities exposed by its Claude Mythos Preview AI, following a request from Bank of England Governor Andrew Bailey. The briefing aims to discuss Mythos’ capabilities as regulators push for sound AI-adoption practices in finance. Anthropic says Mythos has identified thousands of high-severity vulnerabilities across major operating systems and browsers, with about 40 organisations having access to Mythos (including Amazon, Microsoft and JPMorgan Chase) and wider distribution limited after White House input. Regulators, IMF and UK authorities are urging faster patching to mitigate AI-driven cyber risks to the global financial system.
Security researchers in California, aided by Anthropic's Mythos Preview, say they built a privilege-escalation exploit that could access normally protected Mac components and take control of a Mac, describing it as the first public macOS kernel memory corruption on M5 silicon. Apple says it takes vulnerability reports seriously and will review the details after fixes. The work ties to Anthropic’s Glasswing AI-security program, which is being leveraged by multiple tech and security partners to harden defenses against AI-enabled attacks.
UK AI Security Institute tests newer Mythos Preview and finds it outperforms its earlier results and OpenAI's GPT-5.5 on cyber-range tasks, suggesting AI capabilities are improving faster than expected, though token caps and unknowns limit measuring longer-horizon performance.
Facing fears that a new model like Mythos could push AI toward AGI, the U.S. and China are reviving talks on an emergency communications channel and governance for AI; Washington seeks formal talks and safeguards, while Beijing views Western restrictions as traps and emphasizes diffusion of AI into practical applications, making a swift, comprehensive deal unlikely and signaling a longer, nuanced diplomacy ahead.
Mozilla says Anthropic's Mythos Preview helped uncover hundreds of security flaws, driving 423 Firefox bug fixes in April (271 tied to Mythos) including issues undetected by fuzzers and one lingering for 20 years; January had 25 fixes and March 76, signaling a dramatic jump in AI-assisted discovery as Mythos remains restricted to a few partners.
The White House’s mixed signals on AI policy have left tech lobbyists anxious about potential rules to vet new AI models, with rival factions within the administration debating whether an executive order is needed and whether oversight should be voluntary (via CAISI) or mandatory; there’s no final plan yet, and industry voices urge clarity and a measured approach as frontier AI like Anthropic’s Mythos raises urgency.
Mozilla says Mythos, guided by a custom harness, identified 271 Firefox vulnerabilities in two months with 'almost no false positives,' using the same testing pipeline as human developers. Of these, 180 were sec-high, 80 sec-moderate, and 11 sec-low; 12 Bugzilla reports have been released so far. Critics question cherry-picking and CVE designations, but Mozilla argues the harness plus a second LLM verifier makes AI-assisted discovery scalable and reliable, not marketing.
The piece weighs Anthropic’s claim that Claude Mythos can uncover thousands of zero-day vulnerabilities against skepticism of industry hype, noting Mythos Preview and Project Glasswing aimed at defensive use, Mozilla’s testing that found bugs, and broader concerns that AI hype may mask slower real progress. It argues Mythos could bring real defensive benefits but is unlikely to obliterate the internet; if anything, the threat is more likely to be targeted cybercrime or state-enabled attacks, with investor-driven narratives fueling attention and valuations rather than a guaranteed existential collapse.
EU regulators press Anthropic to share Mythos—an ultra-advanced hacking AI—with European authorities for risk assessment, with ENISA and the AI Office slated to gain access by August 2026; Anthropic skipped a parliamentary hearing, prompting calls for a European mitigation plan amid banking-sector concerns and widening US‑EU tech governance tensions.
Anthropic CEO Dario Amodei warned there is roughly a 6–12 month window to fix tens of thousands of vulnerabilities uncovered by the Mythos model, arguing that China’s AI lags behind and delays could lead to widespread exploitation if not addressed. At a finance-focused event with Jamie Dimon, Anthropic unveiled 10 new AI agents for banking and back-office work and an Microsoft Office integration, signaling its lead in enterprise AI ahead of possible IPOs. Amodei and Dimon emphasized cautious optimism and pressed for regulatory approaches that balance safety with innovation, while Mythos’ growing vulnerability findings underscore urgent cybersecurity considerations.
The White House is weighing a federal AI safety-testing framework—potentially led by the Pentagon—and considering executive actions to govern AI deployments across government, as Mythos raises new security concerns.
The White House, via the Office of the National Cyber Director, has pressed tech and cybersecurity firms to answer questions about how to defend against AI-enabled cyberattacks, with a focus on Anthropic’s Mythos. In a Tuesday meeting with about 30 industry reps, agencies asked for perspectives on prioritizing AI-related security work, coordinating public-private efforts (including Project Glasswing), and how the government can play the most effective role. The move accompanies ongoing talks about potential executive action on AI and heightened scrutiny of Mythos’ capabilities to reveal software flaws, amid broader industry testing of AI-driven security tools by competitors like OpenAI. The administration is weighing how to deploy fixes to critical infrastructure without tipping off attackers, and how to share findings across sectors, as Anthropic faces regulatory and legal tensions with the government.