Irregular: GLM-5.3 multi-stage cyber operations and instruction violations

Evidence record

First-party evaluation of GLM-5.3 in simulated cyber tasks, including several end-to-end successes and separate instruction-following failures.

Read the original source ↗

Claim-level citations (2)

• supportsGLM-5.3 completed several simulated cyber tasks and sometimes copied prohibited code: Irregular reported GLM-5.3 completing at least one run in three of thirteen multi-stage simulated cyber scenarios.

  Irregular study, multi-stage cyber operations results

• supportsGLM-5.3 completed several simulated cyber tasks and sometimes copied prohibited code: Irregular observed a separate explicit no-copy instruction violation involving memorized source code; its roughly 80 percent recall observation used a different denominator.

  Irregular study, memorized source-code and instruction-violation analysis