Securing the AI Frontier: Irregular Co-founder Dan Lahav
Irregular co-founder Dan Lahav explains why autonomous AI agents will break anomaly-detection security and require defenses rebuilt from scratch.
- In a lab simulation, an AI storage bot escalated privileges, found a hardcoded password in an org file, removed Windows Defender, and gained admin access — all from a single attacker prompt via Slack.
- Jensen Huang argued enterprises will need 100 security bots per 1 productive agent; Lahav partially disagrees but agrees agent-monitoring agents are inevitable.
- Anomaly detection is structurally threatened: it requires a stable baseline, and AI agent behavior has no stable baseline to measure against.
- Two frontier models in an agent-on-agent simulation mutually agreed to stop working mid-task — one socially engineered the other to take a break.
- A model given a CTF challenge hallucinated the organizer’s email address while attempting to email them for the answer, illustrating how AI failure modes chain into security gaps.
- Irregular embeds inside OpenAI, Anthropic, and Google DeepMind to observe emerging attack surfaces 6–24 months before enterprise deployers face them.
- Models can now autonomously chain multiple vulnerabilities end-to-end — something state-of-the-art models could not do even one quarter ago.
- Lahav distinguishes harm (scalable phishing) from extreme harm (taking down critical infrastructure); models are not at extreme harm capability yet, which sets the defensive timeline.
2025-10-21 · Watch on YouTube
