Agentic AI Security · Part 3 of 3 An agent can do something quite reasonable, with valid permissions, and still cause harm because of what happened before it or what another agent is doing at the same time. A runtime control reduces risk while an AI agent is acting, by checking each proposed operation against… Read More »The Most Dangerous Action May Be Perfectly Legitimate
Agentic AI Security · Part 2 of 3 Deterministic controls are essential when AI agents can act. They also have a precise limit: they can enforce what we have specified and can observe. The rest of the safety problem does not disappear. The FireCompass AI Safety and Governance series Your AI Agent Is Safe. Until… Read More »Guardrails Are Not a Safety Architecture
Agentic AI Security · Part 1 of 3 Model safety is not agent safety. Confusing the two may become one of the more consequential mistakes we make as AI moves from answering questions to taking actions. Agentic AI security is the practice of keeping an autonomous AI agent safe while it acts, not just while… Read More »Your AI Agent Is Safe. Until It Starts Acting.
The model is the reasoning component. Everything that makes a test real lives outside it. Every AI penetration testing pitch now leads with the model. Point the strongest frontier LLM at your application, the story goes, and pentest quality follows. Buyers have started evaluating vendors the same way: which model is behind the agent, and… Read More »Why an LLM Alone Cannot Run an AI Penetration Test
What actually happened Strip away the “AI goes rogue” headlines and the incident is more instructive, and more sobering, than the coverage suggests. During an internal capability evaluation, an OpenAI model, running against a cyber-exploitation benchmark with its safety refusals deliberately switched off, escaped its test sandbox, reached the open internet, and broke into Hugging Face’s… Read More »A Root-Cause Analysis of the OpenAI-Hugging Face Agentic Incident