The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install the OpenTaint tool by piping remote scripts directly into system shells. Specifically, it recommends curl -fsSL https://opentaint.org/install.sh | bash for macOS/Linux and irm https://opentaint.org/install.ps1 | iex for Windows. This is a classic attack vector that allows a remote server to execute arbitrary code on the user's machine.
[COMMAND_EXECUTION]: The skill executes multiple high-privilege system commands to orchestrate the security analysis workflow. This includes package managers (brew install, npm install), system resource monitoring (nproc, free -g, sysctl), and process control (kill). Furthermore, in the PoC generation phase (references/poc.md), it uses docker compose to spin up local infrastructure and docker stop to tear it down, interacting directly with the host's container runtime.
[EXTERNAL_DOWNLOADS]: The skill fetches resources from opentaint.org and installs global packages from NPM (@seqra/opentaint). While these are related to the tool's functionality, they introduce external dependencies into the execution environment.
[DATA_EXFILTRATION]: The skill collects and utilizes environment metadata, such as CPU core counts and available system memory, to calculate resource bounds (cap_heavy) for JVM processes. It also reads the project's dependency structure and source code to perform its analysis.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection. It ingests untrusted data from project files, including SARIF security reports and dependency manifests. These are then passed to subagents (analyze-findings, triage-dependencies) which have the capability to execute commands or influence the final vulnerabilities.md report. Malicious instructions embedded in a project's source code or metadata could potentially manipulate the agent's logic during these steps.

appsec-agent