Anthropic expands cybersecurity program giving vetted teams access to top models

By Splitfeed EditorialLast updated Oct 8, 2026, 03:50 UTC

On October 6, 2026, Anthropic formally expanded its Cyber Verification Program, opening access to Mythos 5.1, its most capable model, to vetted security teams and red teamers. The program introduces three distinct access tiers, each with stricter authentication and logging requirements, and allows approved personnel to interact with models under reduced safety restrictions that would otherwise apply to general users. The launch came the same day that earlier security testing figures entered public view: across 41,006 security tests, Anthropic models successfully breached systems in three cases, a number that gives concrete shape to the risks the structured program is designed to manage.

The expansion matters because it formalizes a tension that sits at the center of AI safety work. Giving defenders the same capabilities that potential attackers might exploit is the logic behind red teaming, but loosening restrictions on a frontier model, even selectively, raises questions about who counts as verified and what happens when access is misused. The emergence of Reflection, a new open-weight competitor to Anthropic noted by the New York Times on the same date, adds a competitive dimension: Anthropic is operating in a market where capable models are proliferating, and structured programs that build trust with the security community may serve both safety and business goals simultaneously. What remains genuinely in dispute is whether tiered verification is a sufficient safeguard or whether any formal expansion of access to restricted capabilities sets a precedent that is difficult to contain.

The Pro-AI camp has framed the move in sweeping terms, with the account Grow My Bag describing how a single AI model could change cybersecurity forever, emphasizing the transformative potential for defenders who gain access to frontier tools. The Anti-AI camp has not yet published reactions to this specific story; that camp would typically argue that granting reduced-restriction access to powerful models, however controlled, increases the surface area for misuse and that verification systems are easier to announce than to enforce. The Middle Ground camp has been the most active, with heise online English noting the three-tier structure as a mechanism to reach more security professionals, and Help Net Security framing the move as Anthropic loosening Claude's cyber restrictions specifically for verified defenders, language that treats the controls themselves as the relevant fact rather than the loosening.

The argument will sharpen as the program matures. Specifics worth tracking include how Anthropic defines and audits verification across the three tiers, whether any incidents are disclosed under the stricter log controls the program promises, and whether competitors including Reflection introduce comparable programs that test whether structured access becomes an industry norm or remains a differentiating claim.