- Google Deepmind announced an “AI Control Roadmap” with zero-trust security protocols for AI agents.
- Google recommends companies with AI agents keep a close watch over them, with plans in place to respond if the agents take harmful actions.
- OpenAI is rolling out new security tools, including a code vetting app and a new GPT-5.5-Cyber model.
Google’s AI lab, DeepMind, has just introduced a new security framework that classifies autonomous, untrusted artificial intelligence agents as internal security concerns, calling them “potential insider threats.”
The roadmap is designed to contain and monitor AI systems that try to bypass human oversight.
At the same time, OpenAI is rolling out a new raft of security features for its own AI tools. In 2026, it seems, top AI companies are taking security risks more seriously than ever.
How Does Google’s New Roadmap Work?
Google Deepmind announced their “AI Control Roadmap” in a blog post that revealed a rough overview of the framework.
It covers “system-level mitigations that limit the harm a misaligned AI system could cause” across four different areas. First is threat modeling, which lists tactics to use against a “hypothetical AI adversary,” assuming a worst-case scenario.
This just in! View the top business tech deals for 2026 👨💻
The other areas covered in the roadmap include two control invariants (specifically, reliable detection of misaligned intent and effective responses to attack attempts), capability-based mitigation ladders (defensive measures that can evolve alongside AI model capabilities), and a portfolio of practical defenses, which is a list of specific, actionable interventions or safeguards.
If you want even more information on all of the above, check out Deepmind’s full 36-page research paper.
Deepmind’s Mitigation Plan
The same announcement also discussed Deepmind’s plan to “deploy AI control mitigations” through a two-pronged approach.
- Identifying the problem (Detection): Uses trusted AI systems to constantly review the less-trusted working agent’s reasoning, actions, and plans in order to keep them on track.
- Managing the problem (Prevention and Response): This gives supervisors their next steps, should they spot an agent taking harmful actions.
While all of these responses might appear to be validating fears of a robot uprising straight out of Terminator or The Matrix, it’s actually a more mundane response.
This new roadmap is essentially taking a zero-trust approach to computer security: We can only be sure that we’re preventing unauthorized system actions by never actually trusting AI agents to be fully autonomous, regardless of how likely they are to go off book with harmful actions.
OpenAI Boost Security Protocols As Well
It’s also worth mentioning that OpenAI has itself announced several security-oriented changes this week.
The company is releasing an updated iteration of the security-focused model GPT-5.5-Cyber and has just released the Codex Security scanner as an app plugin, giving users a simple way to vet code in order to identify, validate, and patch any vulnerabilities.
OpenAI’s ChatGPT and Google’s Gemini are two of the most well-known AI tools operating today.
Seeing both companies working towards increasing security measure surrounding AI agents and tools is an indicator that the cybersecurity ecosystem is well aware of the many challenges that it currently faces as LLMs continue expanding their influence to, well, just about everywhere.
