The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with the Benchling API to retrieve electronic lab notebook (ELN) entries, registry entities, and results. These data sources could potentially contain malicious instructions or jailbreak attempts hidden in metadata, entity names, or notebook text that might influence the agent's downstream behavior.
Ingestion points: Untrusted data enters the agent context via Benchling SDK methods such as benchling.entries.list(), benchling.dna_sequences.get_by_id(), and AWS EventBridge notifications documented in references/eventbridge.md.
Boundary markers: The skill instructions do not specify the use of delimiters or provide instructions for the agent to ignore embedded commands within the retrieved scientific data.
Capability inventory: The agent is granted Bash, Read, Write, and Edit tools, allowing it to perform file system and command-line operations based on the processed content.
Sanitization: There is no evidence of specific input sanitization, filtering, or schema validation for the raw content fetched from the external Benchling API before it is presented to the LLM.
[EXTERNAL_DOWNLOADS]: The skill documentation requires the installation of the official Benchling Python SDK and utilizes the Biopython library. Both are standard, well-known tools in the life sciences and developer community.
Evidence: SKILL.md and references/sdk_reference.md provide instructions to run uv pip install "benchling-sdk==1.25.0" and utilize from Bio import SeqIO for biological file parsing.

benchling-integration