The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements an execute_code function that allows the agent to run arbitrary shell commands using Bash (Linux/Mac) and PowerShell (Windows), as well as Python scripts.
[DYNAMIC_EXECUTION]: The save_skill capability allows the agent to write and persist new Python code as skills, while the execute_code function facilitates the runtime execution of strings as code.
[INDIRECT_PROMPT_INJECTION]: The vision-grounded architecture captures and analyzes screen content and browser data, creating an entry point for malicious instructions embedded in user interfaces or web pages. * Ingestion points: Screen captures via analyze_screenshot, web content via CDPBridge, and local file reading via read_file. * Boundary markers: No specific boundary markers or instructions to ignore embedded content were identified. * Capability inventory: Access to execute_code (shell/python), write_file, and network operations via LLM routing and CDP. * Sanitization: No evidence of sanitization or filtering for perceived visual or text data.
[DATA_EXFILTRATION]: The framework encourages storing sensitive credentials (LLM and framework API keys) in credentials.py or ~/.photoagents/config.json. The agent's file-system tools provide access to these paths, which could be combined with its network capabilities to exfiltrate secrets.
[PERSISTENCE]: The reflection scheduler mode enables the agent to maintain a persistent presence by periodically checking system state and automatically triggering tasks.
[TIME_DELAYED_CONDITIONAL]: The monitoring logic uses conditional triggers based on time elapsed since the last execution (`time.time()
last_run > 86400`) to initiate autonomous behavior.

photo-agents-autonomous-llm