Written By | Torch Media Africa
Anthropic's Mythos and OpenAI's Daybreak can hunt software flaws at machine speed. Now reports say the White House — not the labs — is deciding who is allowed to use them.
Search
TECH & SCIENCE
Imagine an AI that can read millions of lines of code, spot the hidden flaws that hackers dream about, and write the patch before breakfast. Now imagine the same AI in the wrong hands. That double-edged sword is exactly what Silicon Valley's two biggest AI labs have built, and the fight over who gets to hold it has become one of the most important stories in technology today.
In the space of a few months, the answer has shifted from "whoever the lab trusts" to something much bigger. According to a CNBC report on 17 July, the Trump administration is now influencing which companies can access the newest frontier models from Anthropic and OpenAI, a decision the labs once made themselves. The White House says that is not so. The reporting, and the history behind it, suggest the line is blurrier than officials admit.
Two labs, two cyber-weapons, one problem
The story starts with Anthropic's Project Glasswing, built around Claude Mythos Preview, a model Anthropic describes as able to find software vulnerabilities autonomously at scale. Anthropic has kept access tightly restricted, citing safety and national security concerns, and has never made the model commercially available. In April, Mozilla said Mythos had helped it find and patch 271 vulnerabilities in the latest Firefox release — a number that made security engineers sit up.
OpenAI was not going to sit this one out. In April it released GPT-5.4 Cyber, which came with Know-Your-Customer and identity verification requirements. On 7 May it followed up with a trusted-access programme for GPT-5.5, and on 11 May it launched Daybreak, a cyber-defence platform built around GPT-5.5, GPT-5.5-Cyber and its Codex Security agent. Engadget called Daybreak OpenAI's response to Mythos.
Daybreak comes in three tiers. Standard GPT-5.5 handles general enterprise and developer use. GPT-5.5 with Trusted Access for Cyber serves verified defenders doing vulnerability triage, malware analysis, detection engineering and patch validation. And GPT-5.5-Cyber, the most capable and least restricted tier, is in limited preview for authorised red-teaming and penetration testing under controlled conditions, with stronger identity checks and account-level oversight.
OpenAI is careful to say GPT-5.5-Cyber is not mainly a leap in raw capability. It describes it as a more permissive access mode for specialised cyber work, wrapped in verification, monitoring and approved-use limits. That distinction matters, because it means the restriction is about who you are, not what the model can do.
From 'democratise' to 'vetted'
Not everyone liked the gate. As early as 1 May, one AI newsletter ran the headline "OpenAI Quietly Gates GPT-5.5 Cyber After Pledging to Democratize AI Defense", framing the vetted rollout as a retreat from an earlier promise. OpenAI has said it intends to widen access over time. Since then it has expanded Trusted Access for Cyber to thousands of individuals and organisations, and says it plans to widen access to its strongest models through iterative deployment.
Partners in the programme read like a who's-who of the security industry: Cisco, Oracle, CrowdStrike, Palo Alto Networks, Cloudflare, Fortinet, Akamai and Zscaler. From 1 June, individual members using the most permissive models also had to enable Advanced Account Security, because if a model gives real power, a hijacked login becomes a serious weapon.
Enter the White House
Then things got political. Trump's June executive order asked AI companies to give the government early access to models for testing, framed as voluntary. OpenAI said in June that it would limit new models to "trusted partners" to comply with government requests.
The Next Web, summarising CNBC's reporting, said the administration had also blocked Anthropic's Claude Mythos 5 and Fable 5 the previous month over national security concerns, with access restored only after weeks of negotiation. That detail comes from secondary reporting citing two unnamed sources, and Anthropic and the White House have not, to our knowledge, confirmed the account in full.
The centrepiece is a programme called Gold Eagle, launched the same week as an AI clearinghouse for cyber vulnerabilities. According to CNBC's source, Gold Eagle will put the White House in charge of approving which companies can access new models, meaning the partner lists for Glasswing and Daybreak would need explicit government sign-off.
A White House official told CNBC the government does not "provide approvals for AI releases" and that company participation is "voluntary". Critics read it differently. The Next Web's analysis argued that Gold Eagle turns a polite request into a de facto gate, giving Washington distribution power over frontier AI without a new law or a dedicated regulator.
The China factor
The timing could hardly have been more awkward. On the same day as the CNBC report, Moonshot AI launched Kimi K3, a Chinese model that matched or beat Fable and GPT-5.6 on at least one independent benchmark. David Sacks, the former White House AI czar, called that "concerning" and warned that bogging down American labs would cost the US the AI race.
That is the heart of the dilemma. If you hold back your best cyber models, you may stop bad actors — but you may also slow the defenders, and hand ground to rivals who have no such qualms. If you release them freely, you risk arming attackers. There is no option that is both fully safe and fully open.
Europe wants in too. The European Commission is in discussions with OpenAI about access to its models for finding cybersecurity vulnerabilities, according to CyberScoop, a sign that other governments are starting to ask the same question: who decides?
What the experts are saying
Industry voices are cautiously positive about the technology and sharper on the politics. Anthony Grieco, Cisco's chief security and trust officer, calls the technology a "force multiplier for defenders" and argues the value comes from the enterprise framework around the model, not the model alone.
Jared Atkinson, chief technology officer of SpecterOps, says AI speeds up offensive work but does not change the core problem: defenders still need to see attack paths from initial access to critical systems, and visibility into identity exposure is increasingly urgent. Doug Merritt, chairman and CEO of Aviatrix, makes a related point — breach impact depends on how far a compromised identity can move, so containment, not just faster patching, is what limits the damage.
OpenAI chief executive Sam Altman put the optimism bluntly on X: "AI is already good and about to get super good at cybersecurity," adding that OpenAI wants to work with as many companies as possible. On the government side, federal CIO Greg Barbaccia sees potential for Mythos to strengthen federal defences but notes that real-world performance remains uncertain.
Our analysis: every one of these voices is describing a different part of the same elephant. The models are powerful, but they are not magic. The harder questions are about governance — who gets in, who checks, and who is accountable when something goes wrong.
Why 'dual use' is the whole story
Security people have a name for this problem: dual use. A proof-of-concept exploit that confirms a patch works is also a head-start for an attacker. A tool that helps a penetration tester break into a client's network, with permission, is the same tool a criminal would use without it. As one analysis of OpenAI's own rollout noted, separating authorised remediation testing from live exploitation is sometimes a technical question, sometimes a contractual one, and sometimes simply a matter of trusting the user.
That is why account security became a central issue so fast. Once a model meaningfully boosts offensive capability, a stolen password is no longer just a nuisance. Safeguards in the OpenAI programme still block credential theft, stealth, persistence, malware deployment and exploitation of third-party systems, but no filter is perfect.
What this means for you
You will probably never touch GPT-5.5-Cyber or Mythos. But you will feel the results. The banks, hospitals, utilities and apps you rely on all run on software with hidden bugs, and the companies defending them are some of the first customers. Faster patching means fewer breaches and fewer midnight emails telling you to change your password.
The governance fight matters to you too. If a handful of private labs and a single administration decide who gets the best defensive tools, small businesses, hospitals and local governments could end up at the back of the queue. If access is too loose, the same tools could land in criminal hands. Either outcome shapes how safe the internet is for the next decade.
If you run a business, the practical advice from the vendors is consistent: tighten identity controls, enable strong authentication and plan for a world where attackers also have AI. For everyone else, keep your devices updated, because the flaws these models find are exactly the ones that updates close.
The bottom line
A year ago, "who controls frontier AI?" sounded like a seminar question. Today it is a live fight with contracts, executive orders and national security briefings attached. Anthropic and OpenAI built tools that can find the cracks in the digital world faster than any human team. Washington has decided it wants a say in who gets to use them. Beijing's labs are racing to match them.
How this ends will tell us whether the AI era is governed by labs, by governments, or by a messy mix of the two. Right now, nobody — not the companies, not the White House, and certainly not the rest of us — has all the answers.
