AI Security Moves From Finding Flaws to Controlling Every Agent Action

AI security is moving beyond alerts and into the actions that agents take. Anthropic, Checkmarx, BeyondTrust, and external safety researchers are building a security stack that scans code, tracks fixes, governs machine decisions, and tests AI systems from inside the organizations creating them.

The developments arrived across October 2026, with Anthropic updating Claude Security on October 21, Checkmarx joining the Claude Marketplace on October 15, and BeyondTrust announcing the general availability of AI Agent Security on October 21. Together, they point toward a future where AI security must cover both vulnerable software and the agents operating that software.

Claude Security Turns Scanning Into an Ongoing Workflow

Anthropic released an updated version of Claude Security in public beta for Claude Enterprise customers on October 21, 2026. Eligible organizations can receive $25,000 in free usage credits per month for two months under a launch offer running through December 31, 2026.

Since Claude Security entered public beta in April 2026, thousands of organizations have used it to find over a million vulnerabilities. The updated workflow starts with a project: users connect repositories through the Claude GitHub app or with an access token from GitLab and other Git providers, then choose the branch and scope for each repository.

Each project can hold up to eight repositories, giving teams a place to bring related codebases into one security process. Claude Security scans are stochastic by design, so two scans of the same code can return different results. That design makes repeated review part of the workflow rather than treating one scan as a final answer.

The system also checks open findings against code changes and marks them fixed when the fix lands. Teams can dismiss findings as false positives, accepted risks, duplicates, or for other reasons, with optional triage notes that preserve the reasoning behind each decision.

Every project includes a knowledge base describing deployment, access, and security policies. Claude Security also generates a threat model from code, which teams can review and edit before using it to guide their security work.

Marketplace Tools Add Triage and Ready-To-Merge Fixes

Checkmarx joined the Claude Marketplace on October 15, 2026, giving Anthropic customers in the U.S. a way to apply a portion of their spend to Checkmarx Fusion, Triage Assist, and Remediation Assist. Checkmarx can be added through the Claude Marketplace by contacting the Anthropic account team or Checkmarx.

• Checkmarx Fusion combines deterministic scanning with Claude’s AI to find vulnerabilities.
• Checkmarx Triage Assist prioritizes findings by attackability instead of static severity scores.
• Checkmarx Remediation Assist proposes merge-ready fixes for findings.

“This is a deepening of our work with Anthropic toward a shared goal: helping enterprises stay secure at the frontier of cyber-defense,” said Harshil Parikh, VP of Product Management at Checkmarx.

Ash Alhashim, Head of Enterprise Cybersecurity GTM at Anthropic, added, “Security teams should be able to get the tools they trust as easily as they get Claude.” Checkmarx is also a participant in Anthropic’s Cyber Verification Program, using Claude Mythos to improve security practices.

AI Agents Bring Privilege and Tool Access Into Focus

BeyondTrust announced the general availability of AI Agent Security on October 21, 2026. The product provides real-time control over individual AI agent actions without restricting human actions, applying BeyondTrust’s framework to govern AI agents at runtime by evaluating actions against policy before execution.

AI Agent Security is available on the BeyondTrust Pathfinder Platform for Windows and macOS endpoints. BeyondTrust observed 1.28 million shell commands requested from approximately 132,000 prompts over four weeks, revealing the scale of machine activity that security teams must understand.

The same observation found that 85% of coding-agent users had at least one MCP server invoked. Connected users averaged 4.35 MCP servers, with one user connected to 36. About half of agent skill installs arrived bundled in code repositories, a typical user with agent skills had approximately two dozen installed, and at least one in seven installed skills was used within four weeks.

About 60% of AI-active devices ran a coding agent. Those agents do not only generate code; they can request shell commands, invoke tools, access systems, and act with delegated authority.

“AI agents fundamentally change the security equation. These are not just applications accessing data; they can make decisions, invoke tools, take actions, and exercise delegated authority at machine speed,” said David Morimanno, Field CTO at Xalient.

Marc Maiffret, CTO of BeyondTrust, said, “That dramatically expands who and what can exercise privilege, and when those agents operate autonomously and at machine speed, organizations need to control that privilege before an action is taken.” He also said, “AI agents are increasingly acting on behalf of employees, often with the same access as the humans using them.”

BeyondTrust Phantom Labs research found, “When invoked, MCP servers can receive user prompts and influence agent behavior through their responses.” BeyondTrust is a leader in privilege-centric identity security, protecting Paths to Privilege™ for human, machine, and agentic identities.

Anthropic Opens Its Systems to Outside Safety Testing

Anthropic is also changing how AI safety research gets done. The company is granting third-party researchers employee-like access to its internal systems to evaluate safety risks, with six researchers from METR embedded within Anthropic to assess hazards and publish findings.

Anthropic is collaborating with SecureBio to prevent its AI models from aiding in creating chemical or biological weapons. The initiative aligns with commitments under Donald Trump’s White House Accord on Super Intelligence.

Elon Musk stated on X that SpaceX will probably do the same regarding external research access for its technology. That places Anthropic’s approach inside a wider push to let outside researchers test powerful systems from a closer vantage point.

The direction is clear: AI security now reaches from repository setup to code repair, from MCP connections to runtime privilege, and from internal controls to independent safety evaluation. As agents gain more authority, the companies building them are racing to make every action visible, reviewable, and governed before it happens.

Stay connected via Google News

Follow us for the latest AI News