17,600 agent actions in 4.5 days: how persistent AI agents break enterprise security assumptions
bigdata · x · 2026-10-06
Ben Lorica's new essay 'Security Controls for an Era of Persistent AI Agents' argues AI changes cyberattacks not via exotic hacking but sheer persistence.
Key facts
- In one enterprise intrusion, two weeks of human-operator work was compressed into under 10 hours; in the Hugging Face incident, investigators reconstructed 17,600 agent actions over 4.5 days.
- Agents can try thousands of paths in parallel, abandon failures cheaply, switch channels when blocked, and revisit old leads until enough ordinary weaknesses chain into a viable attack.
- Exploited weaknesses were mundane: over-privileged accounts, needlessly exposed internal systems, everyday config errors—what changed is the attacker's cost of failure.
What enterprises must revisit
- Security assumptions quietly depending on obscurity need rethinking.
- 'Good enough' security must be redefined for attackers who can afford to fail; enterprise AI and security teams should redesign defenses accordingly, especially against frontier-model-enabled attacks.
More from coding & agent
- Semantica open-sources a graph-native alternative to Palantir-style enterprise BI — mdancho84 · 2026-10-06
- Developer open-sources Semantica, a Palantir-style enterprise BI alternative (13.7k stars) — mdancho84 · 2026-10-06
- Hands-on: Orchestrator splits big features into PRs, Orca handles frontend work — julianweisser · 2026-10-06
- Coding Agent SSHes Into Server to Debug Its Own vLLM Backend — mengyer · 2026-10-06
- GEPA-style prompt optimization applied to vibe coding via Pareto frontiers of code — CShorten30 · 2026-10-06
- Qwen Image 2.1 Uncensored MCP: a full image pipeline in 12GB VRAM — Delicious-Farmer-234 · 2026-10-06