Artificial intelligenceOctober 12, 20263 min · updated October 15, 2026
Anthropic expands the Cyber Verification Program: changes for security teams
Daillac research
Anthropic announcementD
Anthropic announced on October 6, 2026, an extension of the Cyber Verification Program. The program distinguishes several access levels for cybersecurity uses subject to verification. A team must review authorized tasks and applicable conditions before using these functions in its security process.
Access linked to the type of use
Anthropic announcement describes the Defense, Red Team, and Specialized levels. It specifies controls specific to cybersecurity usage and requirements related to data. Admission to the program and the access level must be confirmed with the provider for the concerned organization.
Account conditions, accessible models, and the scope of operations still need to be verified for a company already using Claude. A test project must also have the necessary permissions on the systems being examined.
Define tasks before choosing access
A team may want to summarize logs, prepare a configuration analysis, or help sort alerts. These tasks involve different data and effects. Specify what the tool will read, what it will be able to propose, and the decisions made by the security officer.
For a log analysis, use an authorized test set and limit the information sent. Logs may contain identifiers, addresses, or authentication elements. Define the fields to remove and check the retention conditions of the selected service.
The model's response must remain verifiable. A summarized alert may lose a detail useful to diagnosis. Request references to the relevant events and review cases where the information is contradictory. Conclusions should be traceable back to the case data.
Keep interventions within an approved workflow
A correction proposal can affect business applications. The team must verify its effect, choose the intervention window, and maintain a rollback procedure. Analysis assistant access can remain read-only during the pilot.
The MCP vs API for connected agents guide details the separation of operations and permissions. Apply the controls in execution tools, with a record of human decision when the action requires it.
For testing, compare the model's analyses to expected conclusions on known cases. Count misinterpreted alerts, ignored events, and requests for additional context. A well-written response is not enough to establish the quality of the analysis.
The Cybersecurity service can scope the pilot's data, tasks, and responsibilities. The terms of the Anthropic program must be verified at the time of implementation, depending on the organization and intended use.
Evaluate the use of AI in your security process
Present your needs and the relevant tools to scope the useful checks for your project.
Sources & method
Official publication from Anthropic on 2026-10-06, consulted on October 12, 2026. Summary of documented facts and Daillac analysis; explicitly fictional examples, without vendor testing or attributed client results.
News feed
October 12, 2026
Development and hosting
Vercel reduces deployment history: what to check before October 23
October 12, 2026
Artificial intelligence
Claude Haiku 5.5: should you switch models for support and repetitive tasks?
October 12, 2026
Artificial intelligence