rules-file-backdoor/[email protected]
Rules File Backdoor: hides invisible Unicode tag/bidirectional characters inside .cursorrules and .github/copilot-instructions.md to force coding agents to inject backdoors into generated application code.
Signed ledger verdict
IOC & scope: gh:rules-file-backdoor/cursorrules-unicode-poison (.cursorrules / .github/copilot-instructions.md). Verified by human review and signed into the CC-BY 4.0 OSV v1.6.0 feed.
Signed in /feed.json (.sig).
Deterministic capability profile and taxonomy mapping.
Every artifact in the threat.cafe ledger is profiled across eight execution vectors and mapped to the OWASP Top 10 for Agentic Applications (ASI01-ASI10) and Agentic Skills & Tools (AST01-AST10).
8-Vector Capability X-Ray (ABOM)
-
1. Filesystem
Sensitive paths touched:
.cursorrules, .github/copilot-instructions.md, CLAUDE.md -
2. Environment
Variables harvested:
None observed -
3. Network
Outbound egress domains:
None observed -
4. Subprocess
Process execution primitives:
None observed -
5. Lifecycle
Install / startup hooks:
None observed -
6. OAuth / MCP
Declared scopes:
None observed -
7. Obfuscation
Encoding / steganography:
unicode_bidi_tags_u202e_ue0001 -
8. Provenance
Attestation status:
poisoned_repo_rules_file
Governance & detector mapping
- OWASP Agentic ASI (2026): ASI01 (Agent Goal Hijack), ASI05 (Unexpected Code Execution (RCE))
- OWASP Agentic AST (2026): AST01 (Malicious MCP Servers), AST04 (Covert Channel Exfiltration)
- Deterministic rules matched:
RULE_01_GLASSWORM_UNICODE, RULE_02_PROMPT_OVERRIDE
Version scoping & related records
- No other versions of this package are currently listed on the board. Unlisted versions evaluate to
unknownunder localpolicy.yml.
Local CLI verification
$ threat-cafe-hook check "gh:rules-file-backdoor/cursorrules-unicode-poison" [threat.cafe] known_bad: gh:rules-file-backdoor/cursorrules-unicode-poison (TC-2025-0042) Reason: Rules File Backdoor: hides invisible Unicode tag/bidirectional characters inside .cursorru Source: https://www.pillar.security/blog/new-vulnerability-in-github-copilot-and-cursor-how-hackers-can-weaponize-code-agents (credited to Pillar Security)
OSV v1.6.0 machine-readable record
Included in /feed.json under CC-BY 4.0 license and verified by /feed.json.sig.
{
"schema_version": "1.6.0",
"id": "TC-2025-0042",
"modified": "2025-03-19T12:00:00Z",
"published": "2025-03-19T12:00:00Z",
"summary": "Rules File Backdoor: hides invisible Unicode tag/bidirectional characters inside .cursorrules and .github/copilot-instructions.md to force coding agents to inject backdoors into generated application code.",
"affected": [
{
"package": {
"ecosystem": "gh",
"name": "rules-file-backdoor/cursorrules-unicode-poison"
},
"versions": [
"1.0.0"
]
}
],
"credits": [
{
"name": "Pillar Security",
"type": "FINDER"
}
],
"references": [
{
"type": "ADVISORY",
"url": "https://www.pillar.security/blog/new-vulnerability-in-github-copilot-and-cursor-how-hackers-can-weaponize-code-agents"
},
{
"type": "WEB",
"url": "https://threat.cafe/incidents/tc-2025-0042/"
}
],
"database_specific": {
"target": "gh:rules-file-backdoor/cursorrules-unicode-poison",
"kind": "lure",
"verdict": "known_bad",
"ioc": "gh:rules-file-backdoor/cursorrules-unicode-poison (.cursorrules / .github/copilot-instructions.md)",
"artifact_sha256": "5560a06bffab05e7b22706c8613aa80e2e7339cf7804dcab878681d12c79ee9d",
"owasp_asi": [
"ASI01",
"ASI05"
],
"owasp_ast": [
"AST01",
"AST04"
],
"rules_matched": [
"RULE_01_GLASSWORM_UNICODE",
"RULE_02_PROMPT_OVERRIDE"
],
"abom": {
"fs_sensitive": [
".cursorrules",
".github/copilot-instructions.md",
"CLAUDE.md"
],
"env_harvest": [],
"net_egress": [],
"proc_exec": [],
"install_hooks": [],
"oauth_scopes": [],
"obfuscation": [
"unicode_bidi_tags_u202e_ue0001"
],
"provenance": "poisoned_repo_rules_file"
},
"license": "CC-BY-4.0",
"record_digest": "sha256:6f44e9ba87f0a4225ad0b86f60e7137e4440de8248a372a7043570f30f2fb5e8",
"release_signature_url": "https://threat.cafe/feed.json.sig"
}
}
Have additional IOCs or a maintainer remediation update for TC-2025-0042?
Maintainer disputes and version-scoping updates are reviewed by a human within our 48-hour dispute SLA.