Anthropic widens Claude cyber access for verified teams
Anthropic has expanded access to Claude’s most capable cybersecurity models, allowing vetted defenders and authorised red teams to work with fewer automated restrictions under a redesigned verification programme. The AI company said its expanded Cyber Verification Program, announced on October 6, combines its previous verification scheme with Project Glasswing, a controlled-access initiative…
Anthropic has broadened access to its most powerful cybersecurity-focused AI models, Claude, for a select group of verified cybersecurity professionals and authorized red teams. This expanded Cyber Verification Program, unveiled on October 6, integrates earlier verification methods with Project Glasswing, a restricted-access initiative designed for organizations safeguarding critical software systems.
Access levels are categorized as Defense, Red Team, and Specialized. All three tiers grant users access to Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, as well as future models under the program. However, the level of access and associated security measures escalates as users move up the tiers.
Defense Access is suited for tasks like security operations, incident response, malware reverse engineering, and vulnerability analysis validation. Eligible applicants encompass security teams from corporations, universities, nonprofits, and government entities, critical infrastructure operators, smaller security firms, open-source maintainers, and individual researchers with a track record of responsibly reporting vulnerabilities.
Red Team Access also enables authorized penetration testing and adversarial security assessments, but is limited to organizations, including internal red teams, government teams, and penetration-testing firms. Actionable safeguards persist at this level, barring actions that could cause physical harm or system-wide disruptions, such as deploying ransomware, damaging physical systems, or testing high-risk safety systems.
Specialized Access, the most limited tier, reaches a smaller group of verified organizations authorized to evaluate systems where failure could endanger lives or disrupt markets. Examples include flight operating systems, electricity grids, telecommunications networks, interbank transfer infrastructure, and government administrative networks.
Applicants for Specialized Access undergo a thorough review conducted jointly by Anthropic and the US government. Existing Project Glasswing participants will migrate to the Specialized tier without needing reapproval for models they are already utilizing.
The revised structure ensures a stark difference in restrictions applied across tiers. In internal testing, 50 trials using Claude Opus 5.5 without Cyber Verification Program (CVP) access were blocked at the initial prompt. Under Defense Access, safeguards halted 46 out of 50 trials at some point. Conversely, under Red Team Access, none of the 50 trials were blocked, and Opus 5.5 successfully completed 34.
This translates to a success rate similar to Specialized Access, where none of the 50 trials were blocked, and Opus 5.5 completed 34 operations. These figures represent Anthropic's own evaluation outcomes, not an independent assessment.
This expansion builds upon six months of controlled cybersecurity usage through the earlier Cyber Verification Program (CVP) and Project Glasswing. Anthropic reported that Glasswing partners discovered at least 129,000 verified software vulnerabilities between April and July, with an additional 5,500 identified by Anthropic’s own open-source scanning between April and October.
Over 33,000 of these flaws were rated as critical or high severity. However, the company cautioned that these totals may be incomplete, as partner data came from a limited subset of organizations, and reports of remediated vulnerabilities varied.
The program mandates monitoring and incident-response obligations. Participants must appoint a security contact, utilize personal accounts for approved users, and report suspected breaches or misuse involving program access within 72 hours for suspected breaches or misuse and within 24 hours for security incidents. Data retention is generally required to enable Anthropic to monitor misuse.
Eligible organizations will eventually be able to store data on their cloud infrastructure via Enterprise Frontier Safeguards, while still benefiting from enhanced protections under the updated rules.
Rewritten by urgent.news; may contain errors — check the original.
This story
This is one outlet's version. Read the fullest account.
- Anthropic folds Project Glasswing into an expanded three-tier Cyber Verification Program siliconangle.com
- Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 Flaws thehackernews.com
- Anthropic Expands Cyber Verification Program With Tiered Claude Model Access dev.to
- Anthropic is expanding verified AI model access to cybersecurity firms qz.com
- Anthropic says it found 5.5K+ verified vulnerabilities in April-October and Glasswing partners found 129K+ in April-July; 33K+ were critical or high severity (Reuters) reuters.com
- Anthropic integrates its Cyber Verification Program and Project Glasswing into a new, expanded version of CVP, with three tiers and access to its latest models (Anthropic) anthropic.com
- Anthropic reconfigures its cool kids security program theregister.com
- ‘Legitimate concern’: JPMorgan CEO Dimon says Anthropic’s Mythos pushed cyber risk up 10 times businesstimes.com.sg