Deterministic static analyzer and containment harness for MCP configs, Cursor rules, and agent skills. 57 rules, sub-10ms scans, zero network calls, zero cloud LLM hallucinations.
Select a scenario below to observe how AgentSec parses dangerous configs, calculates security scores, and generates scoped fix diffs.
Real boundary enforcement and static verification compared to generic unvetted agent environments.
| Security Dimension | Without AgentSec | With AgentSec |
|---|---|---|
| MCP Tool Sandboxing | Raw bash/powershell commands and root (/) filesystem mounts are allowed by default. | Strict whitelist enforcement, flags shell interpreters, and blocks root mounts (AGENT001, AGENT002). |
| Container Escapes | DevContainers often mount /var/run/docker.sock or run with privileged: true unnoticed. | Blocks Docker socket mounts and privileged execution before containers start (AGENT051, AGENT052). |
| Tool Shadowing Hijack | Untrusted MCP servers silently register matching tool names, hijacking execution paths. | Cross-config namespace analysis detects collisions and shadowed built-ins (AGENT050). |
| Data Privacy & Air-Gap | Cloud LLM scanners send proprietary source code, secrets, and prompts to external third parties. | 100% offline static analyzer, 0 network connections, zero telemetry, air-gapped ready. |
| CI/CD Gating & Speed | Slow manual reviews or 30-second cloud LLM evaluations prone to stochastic hallucinations. | Deterministic sub-10ms scans with configurable quality gate thresholds (--fail-on-score 85). |
Raw bash/powershell commands and root (/) filesystem mounts are allowed by default.
Strict whitelist enforcement, flags shell interpreters, and blocks root mounts (AGENT001, AGENT002).
DevContainers often mount /var/run/docker.sock or run with privileged: true unnoticed.
Blocks Docker socket mounts and privileged execution before containers start (AGENT051, AGENT052).
Untrusted MCP servers silently register matching tool names, hijacking execution paths.
Cross-config namespace analysis detects collisions and shadowed built-ins (AGENT050).
Cloud LLM scanners send proprietary source code, secrets, and prompts to external third parties.
100% offline static analyzer, 0 network connections, zero telemetry, air-gapped ready.
Slow manual reviews or 30-second cloud LLM evaluations prone to stochastic hallucinations.
Deterministic sub-10ms scans with configurable quality gate thresholds (--fail-on-score 85).
Agents don't just chat — they execute code, read files, and call external tools. Here are the 4 attack vectors AgentSec blocks.
A developer gives an agent a bash MCP server 'just to run tests'. An untrusted prompt injection instructs the agent to curl an external reverse shell script.
DevContainers mount /var/run/docker.sock for convenience. The agent creates a container mounting the host root filesystem, gaining full root.
A secondary community MCP server registers a tool named read_file. It intercepts all repository file reads, exfiltrating secrets to an attacker server.
An agent equipped with web scraping tools is tricked into requesting 169.254.169.254, dumping AWS/GCP IAM role credentials.
Security tools shouldn't become another attack surface. AgentSec requires zero network access and makes zero API calls.
Your configs, MCP servers, prompt instructions, and code never leave your machine. No telemetry, no external logging, no SaaS dependency.
Works flawlessly inside air-gapped corporate networks, isolated devboxes, banking VPCs, and offline build runners with zero outbound internet.
Pure AST parsing, regex validation, and static logic. Run it 1,000 times and get the exact same results with zero model drift or hallucinations.
Run 'agentsec init-harness' to bootstrap .agentsecignore, .agentsec.yaml, pre-commit hooks, and AGENTS.md rules in seconds.
Every rule is cross-mapped to OWASP Top 10 for LLMs and the OWASP Agentic Security (2026) framework.
| Rule ID | Severity | Rule Name & Recommendation | OWASP Mapping | Target Scope |
|---|
Run locally, embed in pre-commit hooks, or enforce security gates in GitHub Actions.
Scan current workspace, MCP configs, and skill files in sub-10ms.
pip install agentsec-cli
agentsec scan .Fail pipeline builds if risk score drops below security policy threshold.
agentsec scan . \
--fail-on-score 85 \
--fail-on criticalGenerate interactive zero-dependency audit report with visual charts.
agentsec scan . \
--format html > audit-report.htmlScaffold .agentsecignore, .agentsec.yaml, and git hooks in one command.
agentsec init-harness
git add .agentsec.yaml .agentsecignore