The UK AI Safety Institute (AISI) announced that frontier-model cyber-offensive capabilities have crossed a meaningful operational threshold. The findings invalidate the assumptions the defence side has relied on to date.
Striking Findings
Anthropic's Mythos Preview model completed a test that normally takes 20 hours of human red-teaming in 3 out of 10 runs, with 73% success on expert-level tasks. Three weeks later OpenAI's GPT-5.5 showed a similar profile: 2 end-to-end successes out of 10, 71.4% expert-task rate.
Accelerating Pace
AISI estimates that frontier cyber-offensive capability is now doubling roughly every four months — a sharp acceleration from the seven-month pace at end of 2025. The defence side will need a comparable scale-up.
Implications for the Cyber-Security Sector
Static signature- and rule-based security stacks are obsolete against the AI attack loop. XDR platforms like CrowdStrike, Palo Alto and Microsoft Defender have to evolve into AI-native architectures. The defence side urgently needs its own AI capabilities — otherwise the asymmetry will keep widening in the attacker's favour.
Cite this: AI Mevzuları · April 30, 2026 · aimevzulari.com/haberler/siber-guvenlikte-kritik-esik-asildi-ai-saldiri