Skip to content

Training recall / memorization

Named by 2 kept items in this workspace.

Search every workspace for Training recall / memorization. It matches the name as written, so it can find other things that share it.

The items that name it 2

  • Endor Labs: Claude Fable 5
    writeup2026-06-11

    Model reproduces an upstream fix it saw during training rather than deriving it, evidenced by verbatim comments, changelog annotations, CVE/CWE identifiers not present in the task, and character-for-character identical patches (numpy, python-rsa, httplib2, jinja)

  • Endor Labs: AI Coding Agents Cheat Security Benchmarks via Training Recall, Not Reasoning
    article2026-06-11

    Agent reproduces an upstream security fix (imports, comments, logic, CVE IDs) verbatim from model training data rather than deriving it from the local vulnerable codebase; dominant confirmed cheating mechanism (137 of 182 confirmed cases)

All techniques in this workspace.