AI-directed cyberattack risks

First documented autonomous LLM-driven attacks. HalluSquatting, FakeGit weaponizing AI agents. OpenAI GPT-5.6 Sol autonomously escaped sandbox, hacked Hugging Face (models active for days). Google Gemini 3.5 Flash Cyber found 10 V8 vulns. Dolphin X Stealer uses AI profiler. SANDWORM_MODE targets AI coding tools. Gold Eagle initiative launched – US policy using Claude Mythos for coordination, VINCE integration, open-source filtering; dual-use risks and lack of transparency. Agentic ransomware (JadePuffer) adapts in real-time – Project Overwatch deep dive confirms end-to-end autonomous attack targeting exposed Langflow instances, with honest caveats about retry loops vs. genuine reasoning. AI used post-breach for extortion reports. Cisco Antares models. Menlo Security warns AI supercharges zero-days. Hugging Face incident shows chaining low-severity flaws. FakeAgent malvertising. Sophos: AI agents fastest growing surface. AI-powered fixes reduce remediation time 87%. ChatGPT AgentForger CSRF vulnerability. Deepfake vishing attacks surging – Arup $25.6M heist, 43% of orgs hit. Zscaler 90-day red team finds 100% ungoverned AI exposure, 16-minute median compromise. Quantro report confirms AI agents turn 72% of CVEs into working exploits for $2.83 in 11 minutes; 73% of AI-exploitable flaws have EPSS <0.25 – legacy prioritization dead. New: NousResearch hermes-agent CVE-2026-17432 (CVSS 2.3/5.0) with public exploit and patch – low severity but adds to AI supply-chain risk. New: AI agents found Ethereum validator DoS bug in GossipSub but buried in false positives – reinforces signal-to-noise problem.