The Agent Skills Directory

[DYNAMIC_EXECUTION]: The example script examples/interleaved-thinking/examples/03_full_optimization.py implements a mock calculator tool using the Python eval() function. Although it attempts to sandbox the execution by restricting __builtins__ and using a whitelist of math functions, eval() is notoriously difficult to secure in Python and can often be bypassed to achieve arbitrary code execution.
[PERSISTENCE]: The repository includes a 'Continuous loop' feature for its researcher component, which uses macOS launchd service definitions (located in researcher/orchestration/launchd/) to automatically run scripts like loop_step.py and loop_daily.py at regular intervals or on startup. This is intended functionality for the 'Researcher Operating System' described in the project documentation.
[EXTERNAL_DOWNLOADS]: The script examples/long-horizon-prompt-lab/scripts/fetch_report.sh downloads PDF files from cdn.openai.com. These are used as research artifacts for the long-horizon prompting skill. While openai.com is a well-known service, the skill's continuous loop also fetches content from arbitrary URLs provided in the research queue.
[INDIRECT_PROMPT_INJECTION]: The 'Researcher' and 'Evaluation' components (e.g., examples/llm-as-judge-skills/src/tools/evaluation/direct-score.ts and researcher/scripts/loop_step.py) ingest external, untrusted data from URLs or files and interpolate them directly into LLM prompts for analysis, scoring, or synthesis. This creates a significant surface for indirect prompt injection where malicious instructions embedded in research sources could influence the agent's behavior during the research loop.

context-engineering-collection