Cybersecurity AI now lives in gated communities

Welcome to Runtime! Today: New models from Anthropic and OpenAI claim impressive and scary cybersecurity capabilities, but only club members know for sure, NIST explains how it will deal with a huge surge in vulnerability reports, and the latest enterprise moves.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week, and if you value independent enterprise tech journalism, click the button below and become a Runtime supporter today.

Role-based access

Over the last six months, rapid advances in AI models built by Anthropic and OpenAI have forever altered the software development process and opened the door to broader adoption of AI agents throughout the enterprise. However, that progress comes at a cost, now that those models have the potential to cause new chaos for security teams.

Anthropic released Opus 4.7 Tuesday, the first new version of its flagship model since the release of Opus 4.6 in November last year kicked off a new chapter in enterprise AI. The company claimed the new model is even better at agentic coding than its predecessor, but cybersecurity professionals who want to use it will need a hall monitor.

• Opus 4.7 is the first model for which Anthropic will implement "safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses," the company said in a blog post, which follows last week's announcement of Project Glasswing.
• Dozens of companies and organizations were invited to sample the too-dangerous-for-public-consumption Mythos Preview model, and Opus 4.7 was actually designed to limit those capabilities as a test of the safeguards.
• "Security professionals who wish to use Opus 4.7 for legitimate cybersecurity purposes (such as vulnerability research, penetration testing, and red-teaming) are invited to join our new Cyber Verification Program," the company said.

OpenAI introduced a similar program earlier this year called Trusted Access for Cyber, and expanded that program this week alongside the release of GPT-5.4-Cyber, "a variant of GPT‑5.4 trained to be cyber-permissive." Much like Anthropic's strategy with Mythos Preview and Opus 4.7, GPT-5.4-Cyber will only be released to "customers in the highest tier" of that program, it said in a blog post Tuesday.

• The company announced the first batch of customers in that highest tier on Thursday, and it includes many of the same organizations as Project Glasswing, including CrowdStrike, JPMorganChase, and Nvidia.
• Interestingly, however, OpenAI did not name any of the major cloud providers as part of that tier, except for major-ish cloud provider Oracle, which has a rather special relationship with OpenAI.
• "We want participants to push the frontier of defensive research, share what they discover, and help turn new insights into stronger protection for everyone," OpenAI said Thursday.

Enterprise AI has moved so far, so quickly in just a few years, and a collective decision to take the industry's foot off the gas pedal in the name of security is probably a good thing. But the downside is Anthropic and OpenAI just set up a cybersecurity caste system that will entrench the power of hyperscalers and other massive security vendors while promising to extend the benefits to the rest of enterprise tech, and it's not clear how that will actually happen.

• The history of enterprise software seems to always come back to playing on the fear, uncertainty, and doubt of tech buyers when it comes to the power of new technologies, so in some way Anthropic and OpenAI's cybersecurity strategies are just a new link in the chain.
• And these programs could give the participants time to understand how future models will cause cybersecurity issues, which presumably also gives them time to act to prevent those issues from ever happening.
• But it feels like they're really giving the current generation of power brokers in enterprise tech a look behind the curtain that challengers to that throne are simply unable to see.
• And that means that the future cybersecurity products built by vendors with the knowledge they gained inside the clubhouse will have a huge head start, unless Anthropic and OpenAI make a concerted effort to bring startups and smaller vendors into the fold.

Holes in the bug screen

The beleaguered Cybersecurity Vulnerabilities and Exposures (CVE) program was struggling for months amid uncertainties about the future of its funding sources, and a sharp rise in AI-driven reports of new vulnerabilities pushed it to a breaking point this week. The National Institute for Standards and Technology (NIST), which administers one of the most important tools in enterprise cybersecurity, announced Thursday that the CVE program would no longer add key details about new vulnerabilities unless they rise to a certain level.

As of Wednesday, NIST will no longer "enrich," or add details such as a severity score and products affected, to all vulnerabilities discovered in the future unless they meet certain criteria. The ones that will still get those crucial details include vulnerabilities that CISA knows are currently being exploited, vulnerabilities in software used by the federal government, and vulnerabilities as described by a Biden-administration project to define "critical" software.

NIST cited a 263% increase in vulnerability submissions over the last five years as a reason behind its decision, and said that trend continued into 2026. "They’ve just come out and publicly stated, 'We are never going to get through this backlog,'" Trend Micro's Dustin Childs told CSO, which is a backlog that could be addressed — or at least reduced — with proper federal funding.

Enterprise moves

Duane O’Brien is the new executive director of the Open Source Initiative, joining the guardian of open-source licensing after technology leadership roles at Capital One and Indeed.com.

Anuj Kumar is the new chief revenue officer at Backblaze, joining the cloud storage company after leadership roles at SUSE, Human Security, and NetApp.

The Runtime roundup

U.S. data center operators will soon be required to submit power consumption data to the Energy Information Administration, which is part of the Department of Energy, according to Wired.

Google Cloud struck a deal with private equity giant Thoma Bravo to provide access to Gemini models for companies in its portfolio, which contains a lot of SaaS companies.

Thanks for reading — see you Saturday!