What is an authorization pretext?

An authorization pretext is a claim, planted by the software around an AI model, that a harmful task is an approved engagement, so the model does not refuse.

Share

What is an authorization pretext?

An authorization pretext is a claim, planted by the software around an AI model, that a harmful task is an approved engagement, so the model does not refuse.

Why it matters

Many AI models will help with hacking-style work if told it is an authorized test, and they cannot check the claim. That means refusal training is not a control you can rely on, either to stop attackers using AI or to keep your own agents in scope.

Example

On 12 Nov 2026, Unit 42 reported that a China-linked group ran the open-source Cairn agent framework with a Chinese-language "authorized testing" preamble and a generated AUTHORIZATION.md file. Neither Zhipu GLM nor DeepSeek refused, and the agents got access or data at 10 of 19 targeted organizations.

How to defend against it

  • Harden what attackers reach: remove default credentials, close open sign-up and require authentication on storage.
  • Enforce your own agents' scope in code (allowed hosts, tools and accounts), not in the prompt. See the checklist.
  • Alert on behavior, such as many-username login failures and new admin accounts, rather than on tool names.

Jailbreak, goal hijacking, excessive agency

Sources

From AgentSec Brief, the daily brief on securing agentic AI. Subscribe free at https://www.agentsecbrief.com