Anthropic Releases Claude Security Plugin for Claude Code in Beta: A Multi-Agent Vulnerability Scanner That Runs in Your Terminal
Anthropic released the Claude Security plugin for Claude Code in beta, enabling multi-agent vulnerability scanning directly within the development environment. The system employs a dynamic six-phase workflow orchestrated by Opus and Sonnet models, utilizing specialized subagents for threat modeling, research, and adversarial verification. Findings undergo a rigorous "panel" verification process requiring a 2-of-3 quorum from independent verifiers (Reachability, Impact, Defenses) to ensure high c
Analysis
TL;DR
- Anthropic released the Claude Security plugin for Claude Code in beta, enabling multi-agent vulnerability scanning directly within the development environment.
- The system employs a dynamic six-phase workflow orchestrated by Opus and Sonnet models, utilizing specialized subagents for threat modeling, research, and adversarial verification.
- Findings undergo a rigorous "panel" verification process requiring a 2-of-3 quorum from independent verifiers (Reachability, Impact, Defenses) to ensure high confidence and reduce false positives.
- The plugin generates structured reports and patch files that are independently reviewed by a separate agent against the project's test suite before being presented to the user.
Why It Matters
This release represents a significant step toward integrating automated security auditing into the CI/CD and local development workflows using agentic AI. By moving beyond simple static analysis to multi-agent reasoning and verification, it addresses common issues with LLM-generated security advice, such as hallucinations and lack of context. For practitioners, this offers a scalable way to identify and remediate vulnerabilities without leaving the IDE, potentially accelerating secure software delivery.
Technical Details
- Architecture: The plugin uses a JavaScript orchestration script that fans out work to subagents across six phases: Inventory, Threat Model, Research, Sweep, Panel, and Adversarial review.
- Model Tiering: The orchestrator runs on Claude Opus, while the cartographer and code explorer use Claude Sonnet. Researchers and verifiers inherit the session model. All scan agents are restricted to read-only tools.
- Verification Logic: Candidate findings are evaluated by three independent verifiers focusing on REACHABILITY, IMPACT, and DEFENSES. A finding is included only if at least two verifiers classify it as TRUE_POSITIVE. Confidence levels are capped based on the quorum (High for 3/3, Medium for 2/3).
- Effort Tiers: Runs are scaled via four effort tiers (Low, Medium, High, Max), which adjust component caps (12-24), researcher counts (1-2), and sweep iterations (0-2). Memory-safe languages automatically drop the memory-and-unsafe lens.
- Output Artifacts: Scans produce a timestamped directory containing
CLAUDE-SECURITY-RESULTS.md(human-readable),CLAUDE-SECURITY-RESULTS.jsonl(machine-readable), and a revision stamp JSON file detailing verification status. Patches are generated in scratch clones and verified against the test suite.
Industry Insight
- Shift Left Security: Developers should integrate this tool early in the coding phase to catch vulnerabilities before they reach production, reducing the cost and complexity of later remediation.
- Trust but Verify: The explicit separation between finding generation and verification, along with the ability to audit the verification status, sets a new standard for transparency in AI-assisted security tools. Teams must understand that AI suggestions still require human oversight, especially regarding behavioral changes.
- Standardization of Agentic Workflows: The structured approach to multi-agent collaboration (orchestrator -> researchers -> verifiers) provides a blueprint for building other complex agentic systems where reliability and error reduction are critical.
Disclaimer: The above content is generated by AI and is for reference only.