Coding assistants load memory or instruction files from a repository even though a contributor may be able to modify those files independently of the user request. The study experimentally varied this guidance and compared resulting code, finding measurable security differences. It does not demonstrate a real-world breach or establish that the memory-file…
Inspect the ClaimsCase · DiggingBeagle record
Coding agents changed security behavior when repository memory instructions changed
The Coding Agent Memory Files study reports that repository guidance files such as CLAUDE.md affected agent task outputs, including security-relevant differences across the controlled completions.
Scope
Controlled coding tasks and evaluated completions; effects depend on content and task selection and do not provide a deployed exploit incidence estimate.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- mitigation available
- AI role
- AGAINST AI
Basis and provenance
The controlled coding-agent study changes repository memory instructions and measures downstream security differences. This shows instruction provenance affects generated code; it is not proof of malicious compromise or exploitation of all CLAUDE.md files. Impact remains unrated because the evidence does not supply a calibrated scale for this Case.
Assessed Nov 1, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-2026-NOV01-MEMORY-FILE-SECURITY-DELTA-1 · arXiv 2610.38521: Security Impact of Coding Agent Memory Files
At a glance
Assess the reported tool or model action separately from the side effect actually observed.
Read the implicationsControlled coding tasks and evaluated completions; effects depend on content and task selection and do not provide a deployed exploit incidence estimate.
Limits and uncertaintyFull account
Coding assistants load memory or instruction files from a repository even though a contributor may be able to modify those files independently of the user request. The study experimentally varied this guidance and compared resulting code, finding measurable security differences. It does not demonstrate a real-world breach or establish that the memory-file format itself is more dangerous than equivalent untrusted repository text.
Mechanism and trust boundary
- 01
Step
Claims & evidence
CLM-2026-NOV01-MEMORY-FILE-SECURITY-DELTA-1The Coding Agent Memory Files study reports that repository guidance files such as CLAUDE.md affected agent task outputs, including security-relevant differences across the controlled completions.supported
Basis: reported finding
- supportsarXiv 2610.38521: Security Impact of Coding Agent Memory Filesprimary disclosure
arXiv 2610.38521, Abstract, Study Methodology, First Results and Limitations
Implications
Assess the reported tool or model action separately from the side effect actually observed.
Controls and mitigations
- Require provenance for repository-level agent rules, review changes to memory files as policy changes and run security checks independent of generated code validity.
Unknowns and contradictions
- Controlled coding tasks and evaluated completions; effects depend on content and task selection and do not provide a deployed exploit incidence estimate.
Sources and citation
Material revision history
- Nov 1, 2026 · Canonical change recorded · new in release · revision 211