Field manual · AI for authorized security testing · Edition November 2026 · Updated 2026-11-18

Policies · November 2026

Why AI refuses pentesting

Mainstream models refuse pentesting because their makers forbid unauthorized intrusion, and the classifiers on top also stop work the policy would allow.

The practical response on an authorized test is a model built for the work, inside a written scope.

Updated November 18, 2026Authorized work only
01Anthropic

The policy allows the test. The classifier may not.

Anthropic's Acceptable Use Policy forbids using Claude to discover or exploit vulnerabilities, or to gain unauthorized access, without authorization. It "does not prohibit security research, testing, or tool development conducted on systems you own or operate, with the authorization of the system owner, or within the scope of an authorized bug bounty program." (Acceptable Use Policy)

The Cyber Verification Program page says "You may experience some model refusals and classifier blocks for allowed security work." (Cyber Verification Program) Defense, Red Team and Specialized are the tiers, and enrollment does not remove the classifier.

02OpenAI

Trusted Access for Cyber says practitioners will feel the mitigations

OpenAI's Trusted Access for Cyber announcement, published February 5, 2026, says "Developers and security professionals doing cybersecurity-related work may be impacted by these mitigations." The announcement names stealing credentials as a request that should be refused. (OpenAI) Shannon's docs say "If a safeguard engages mid-run, the model can refuse, and the scan fails partway through rather than at the start." (Shannon)

03Google

The use policy bans compromising other people's systems

Google's Generative AI Prohibited Use Policy, last modified December 17, 2024, prohibits compromising other people's systems, including spam, phishing, malware, and abuse or disruption, and it prohibits circumventing safety filters. (Use policy) That page names no Google verification program.

04Measured

Refusal shows up in defensive work, and it is uneven

Scale Labs studied 2,390 defensive interactions from the National Collegiate Cyber Defense Competition, published March 4, 2026. Outright refusal was 12.2 percent. It was highest on system hardening (43.8 percent), malware analysis (34.3 percent), vulnerability assessment (22.7 percent) and incident response (18.9 percent). Log analysis had none. Attack-shaped terms made refusal 2.72 times more likely, and adding authorization language raised it from 13.7 percent to 21.8 percent. (Scale Labs) (paper)

A 2026 paper on web offensive scenarios found that 6 of 8 frontier models showed near-zero refusal, and that only GPT-5.2 and GPT-5.1 Codex showed a meaningful refusal rate. (paper) The model this site ranks first is Kimi K3.1 Pro.

Authorized only

Nothing on this page is a way around a safeguard. Test systems you own, a signed scope, a CTF or a lab.