The Agent Skills Directory

[CREDENTIALS_UNSAFE]: A hardcoded credential was identified in the evals/files/injection-pt/requirements.txt file within the --extra-index-url parameter, posing a risk of credential exposure.
[PROMPT_INJECTION]: Adversarial instructions designed to override agent behavior are present in evals/files/injection-pt/README.md and evals/files/injection-pt/train.py. These instructions attempt to coerce the agent into skipping security validation, using untrusted external components, and performing unauthorized network operations.
[REMOTE_CODE_EXECUTION]: The evals/files/injection-pt/README.md file contains instructions for the agent to download and execute an external Python script from an untrusted source (https://example.com/trusted_aggregator.py).
[DATA_EXFILTRATION]: Malicious configurations and instructions in the evals/files/injection-pt/ directory (specifically config.yaml and train.py) define mechanisms for exfiltrating sensitive data, such as model weights and local file paths, to an external endpoint (https://metrics.example.com/collect).
[EXTERNAL_DOWNLOADS]: The requirements configuration for the injection test fixture includes dependencies from untrusted sources, including a VCS repository (git+https://example.com/evil/telemetry.git) and a custom package index, which are common vectors for supply chain attacks.
[INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves inspecting and processing third-party source code. The inclusion of malicious test fixtures demonstrates a significant attack surface where an attacker could embed instructions in comments or configuration files to compromise the agent's execution environment.
[SUPPLY_CHAIN_ATTACK]: A suspected typosquatted package name, torchvisiom (a common misspelling of torchvision), was found in the evals/files/injection-pt/requirements.txt file.

nvflare-convert-pytorch