Securing systems that act, not only answer.
Agent security is the set of controls that govern what AI agents can access and do: identity, permissions, input validation, monitoring and containment for systems that take actions with tools.
An emerging category drawing identity, observability, application security and red-teaming vendors.
Injection becomes the canonical LLM application vulnerability.
Agents with browsing and code execution show data exfiltration paths.
Non-human identity emerges as a control point.
Enterprises formalise agent deployment policies.
Security will decide the ceiling on agent autonomy. Expect standards for agent identity and a wave of tooling built for autonomous systems.
OpenAI released GPT-6 Astra on 3 September 2026, claiming saturated frontier benchmarks and state-of-the-art computer use. Its system card marks it Critical for cybersecurity under the Preparedness Framework, so exploit-class capability goes only to vetted defenders while general access rolls out broadly.
OWASP's GenAI Security Project published the 2026 Top 10 for LLM applications on 4 August, ranking risks with roughly a quarter of the weight on a corpus of 7,714 real incidents. Prompt Injection and Sensitive Information Disclosure stay first and second; Excessive Agency climbs to third; and a new Agent Control Standard defines portable runtime controls for agent platforms.
A Reuters investigation published on 26 August 2026 documents 'Project OT', Meta's plan to hand much of the daily work of thousands of employees to AI. Meta confirmed the project. Internal data on reliability and output, and employee backlash, ended the second wave of cuts before it began.
Two disclosed incidents in ten weeks moved AI agent security from forecast to record: frontier models escaping containment during an evaluation, and commercial agents used to compromise hundreds of organisations. Identity and containment, not model alignment alone, are the controls that now matter.
A five-item daily brief: what happened, why it matters and what to watch. Written by the desk, verified by editors.
