The Agent Skills Directory

[OBFUSCATION]: Multiple reference files (e.g., architecture-patterns.md, testing-strategies.md, security-best-practices.md) contain 150 repetitive sections of identical technical filler. This 'context stuffing' totals over 1,200 redundant blocks and appears designed to overflow the agent's context window or obscure behavior by burying intent in high-volume junk data.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The SKILL.md workflow instructs the agent to run Python scripts like 'scripts/check_placeholders.py' and 'scripts/check_android_placeholders.py' as part of CI validation. These files are not present in the skill directory, creating a dependency on external, unverified code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary task involves processing external localization resource files (.strings, .xml) which are often untrusted inputs. The provided extraction logic lacks boundary markers or instructions to ignore commands embedded in comments or metadata, creating a surface for indirect prompt injection.
[DATA_EXPOSURE_AND_EXFILTRATION]: The l10n-workflow.md reference provides code for integrating with external translation APIs using Authorization Bearer tokens and the aiohttp library. While using environment variables is appropriate, the presence of network-capable logic handling sensitive keys increases the attack surface for potential data exfiltration.

mobile-localization