AI Cybersecurity Models Pose Escalating Global Security Risk
Anthropic's Mythos AI model discovered previously undetected software vulnerabilities, prompting the US government to restrict its deployment to trusted organisations. OpenAI and Anthropic disagreed publicly over whether to grant European regulators access to their cyber AI tools, raising questions about governance of systems capable of large-scale vulnerability exploitation.
What happened
Anthropic's Mythos AI model, unveiled in April 2026, discovered previously undetected vulnerabilities in major operating systems and browsers, some dormant for decades. The US government issued a national security directive suspending public access and restricting deployment to 'trusted' American organisations.
OpenAI and Anthropic publicly disagreed over whether their new AI cyber tools should be shared with European regulators. OpenAI offered Brussels access to its model while Anthropic withheld theirs, describing negotiations with the European Commission as being at a different stage.
Reported impact
AI models capable of discovering and weaponising software vulnerabilities at unprecedented scale pose potential for automated cyber warfare targeting critical infrastructure including power grids, financial networks, satellite communications and military systems. Such conflicts could escalate and conclude at machine speed, beyond human observation and decision-making capacity.
Index analysis
The development mirrors concerns about AI-powered defensive capabilities being necessary to counter offensive AI attacks. China reportedly created a comparable 'cyber nuclear weapon' with equivalent capabilities to Mythos, raising concerns about an escalating global AI arms race.
Governance controls
The US government imposed national security restrictions on Mythos deployment and redeployed access only to restricted trusted organisations. Regulatory disagreement between OpenAI and Anthropic over European regulator access to cyber AI tools remains unresolved, raising questions about appropriate governance mechanisms for dual-use AI systems.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.