The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions in HEARTBEAT.md and references/security-patterns.md to scan for and ignore injection attempts such as 'ignore previous instructions'. These are purely defensive measures designed to protect the agent from external adversarial inputs.
[COMMAND_EXECUTION]: The skill provides scripts/security-audit.sh, a bash script that performs local security checks such as verifying file permissions on credentials and scanning for exposed secrets. These operations are benign and follow security best practices.
[DATA_EXPOSURE]: The skill establishes clear rules for credential management, advising that secrets be stored in a .credentials/ directory and including a script (security-audit.sh) to verify permissions and scan for accidental exposure. It also mandates that these files be excluded from version control via .gitignore.
[INDIRECT_PROMPT_INJECTION]: The architecture is designed to handle untrusted data by defining a clear separation between user instructions and external content (websites, emails). It mandates that external content be treated as data to be analyzed rather than commands to be executed, effectively mitigating indirect prompt injection risks.
[PROMPT_INJECTION]: The skill defines a 'proactive' identity for the agent, using directives such as 'Relentless Resourcefulness' and 'Ask forgiveness, not permission'. These instructions are intended to increase the agent's autonomy within the workspace and are constrained by a 'Nothing external without approval' rule to prevent unintended external impacts or security breaches.

proactive-agent