Notifications
Clear all
Topic starter
18/08/2026 2:54 pm
TL;DR: GPT-5.5 is the first OpenAI model rated High for cybersecurity, and MindFort reports that its real-world pentesting performance improves materially inside an autonomous agent harness while still falling short of verifier-confirmed end-to-end exploit chains. The governance lesson is that model capability now depends as much on orchestration, tool access, and validation loops as on the base model itself.
NHIMG editorial — based on content published by MindFort: How Good Is GPT-5.5 for Cybersecurity?
By the numbers:
- GPT-5.5 scored 41 in MindFort's NexBench evaluation, with 45 findings validated against a live target.
- CrowdStrike reported an 89% increase in attacks by AI-enabled adversaries in 2025 versus the previous year.
Questions worth separating out
Q: How should security teams govern AI agents that can choose tools at runtime?
A: Security teams should govern runtime agent choice as an access event, not as a simple application action.
Q: Why do AI SOC agents create a new access-control problem?
A: Because they need credentials and permissions to query multiple security tools, but they also make runtime decisions that traditional scripts cannot.
Q: What do organisations get wrong about AI-driven cyber risk?
A: They often assume the main change is autonomous attackers, when the immediate change is faster and more variable abuse of existing identity pathways.
Practitioner guidance
- Define agent identities explicitly Register every AI agent, orchestration service, and cyber-capable model under a distinct identity with documented owner, purpose, and approval scope.
- Constrain tool access by task Limit each agent to the smallest connector set needed for the workflow, then require step-up approval before it can reach code execution, production data, or external systems.
- Require verifier-backed validation Treat model findings as hypotheses until a second control confirms them through independent reproduction, policy checks, or sandbox validation.
What's in the full article
MindFort's full article covers the operational detail this post intentionally leaves for the source:
- NexBench scoring methodology and the validator rules used to accept or reject findings
- Per-model benchmark tables showing effort level, run time, accepted findings, and cost
- The GPT-5.5-Cyber access model, including what is allowed, what is refused, and who gets gated access
- Comparisons with OpenAI, UK AISI, and other external evaluations of cyber capability
👉 Read MindFort's analysis of GPT-5.5 cybersecurity capability and agentic risk →
GPT-5.5 and agentic cyber risk: are your controls keeping up?
Explore further
18/08/2026 3:37 pm
AI security has crossed into identity governance because the model now operates as a controllable actor. Once a model is wrapped in tools, memory, and orchestration, the question becomes who or what it is allowed to be at runtime. That is not just model safety, it is identity and privilege design for software that can act. NIST AI RMF and OWASP-AGENTIC both point toward this shift, but the operational control plane will look increasingly like IAM and PAM for agents. Practitioners should stop treating model access as a generic API problem.
A question worth separating out:
Q: Who is accountable when an AI agent exceeds its intended scope?
A: Accountability should follow the delegation chain, not stop at the agent label. The human requester, the policy owner, and the team that granted underlying access all matter, because the agent acts within a permission model someone designed. If the chain is unclear, the governance model is already too weak.
👉 Read our full editorial: GPT-5.5 shows why autonomous cyber models need runtime controls
