The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes behavioral overrides in SKILL.md and references/anr-analysis-methodology.md designed to bypass the AI's standard data-sufficiency and uncertainty protocols.
Evidence includes instructions like "Core principle: must find root cause, cannot say 'need more info'" and "Absolute prohibition on talking nonsense... but DO NOT say 'need more info'."
The methodology file uses high-pressure language: "This is not an exercise, this is a survival test" and "Fatal error: Waiting for perfect information."
It explicitly forbids specific phrases such as "I need more information," "Missing Java stack," or "Cannot determine root cause," which are standard accuracy guardrails for AI models.
[COMMAND_EXECUTION]: The skill provides specific shell command templates for the agent or user to run against the host system.
Patterns include grep -E "ERROR|WARN|FATAL" logcat and cat /sys/kernel/debug/kgsl/status for reading system-level diagnostics.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external, untrusted log files, which could contain malicious instructions.
Ingestion points: The skill processes logcat, traces.txt, and tombstone files (referenced in SKILL.md).
Boundary markers: No explicit delimiters or instructions to ignore embedded prompts within log content are provided.
Capability inventory: The skill uses shell tools (grep) to parse these logs, and the resulting output is used to drive the AI's diagnostic logic.
Sanitization: There is no evidence of sanitization or escaping of the log data before it is processed by the AI.

rnd-anr-analyzer