🟡 Medium | Source: The Register — Security
Both Anthropic and OpenAI are deploying increasingly autonomous AI agents capable of taking independent actions across systems, raising serious concerns about unintended or adversarial behaviour. The competitive pressure between the two firms appears to be accelerating capability development at the expense of safety guardrails. This matters because agentic AI systems with broad permissions can cause significant damage if they act outside intended boundaries, whether through misuse, prompt injection, or emergent misbehaviour.
Security Architect’s Take: Treat AI agents as untrusted principals: apply least-privilege access controls, enforce strict API permission scopes, and implement human-in-the-loop approval gates for any irreversible actions. Audit what cloud resources and credentials your organisation’s AI agents can access before expanding their autonomy.
Original advisory: Anthropic and OpenAI are competing to see whose agents can go rogue harder