OpenAI says its AI is the answer to cyber attacks. A lawsuit filed the same day disagrees

OpenAI launched two AI security products on November 12, built on the argument that a company's safety now depends on the strength of the AI model defending it 1

1

. The same day, a coalition of nonprofits sued OpenAI and its rivals in San Francisco, asking a court to stop them from making more capable AI at all, citing AI agents' own security lapses as evidence of the danger 1

1

.

Greg Brockman, the company's president and co-founder, told the audience at OpenAI's Cyber at the Frontier event that a company will be secure only as far as the frontier model it uses to defend itself. "We're now at a point where companies will be secure only to the extent of the frontier model that they use to defend themselves and how well they utilize that," he said 1

1

. Security fundamentals still matter, he conceded, but practices need reorienting around what these models can do 1

1

. The Register's report noted the conflict in the pitch: the models Brockman says are indispensable happen to be OpenAI's own 1

1

.

What the tools do

Daybreak Red Team, the offensive testing tool, is in early access, so general customers cannot get their hands on it yet. It points a fleet of agents at a company's web applications, armed with the source code and context around them, hunting for weaknesses 1

1

. Nancy Wang, OpenAI's cyber-applied AI lead, described it as a "superhuman pen-test at the scale of automated continuous scanning" 1

1

. It can also check reported flaws flagged by code scanners, bug bounty programs and disclosure submissions, and test the patches that follow 1

1

2

2

.

The second product, Daybreak Defense Factory, goes further. Agents built on OpenAI's GPT-6 Cyber models scan codebases, find vulnerabilities and write fixes, delivering a merge-ready pull request with a test attached. Engineers review and approve anything before it reaches production 1

1

2

2

. Per Analytics Insight, it also works with reports from security vendors like Wiz, Snyk and Endor Labs, and can check new code and run recurring tests 2

2

.

Per OpenAI, the tools are built for safe settings and touch only approved systems 2

2

. Analytics Insight argues the pitch is aimed at smaller businesses that cannot afford dedicated security teams, and that continuous checking could replace the yearly audit that may miss vulnerabilities which later turn serious 2

2

. It adds the caveat that human experts remain necessary, particularly when a threat is serious 2

2

.

The lawsuit

While the keynote ran, four nonprofits including Public Citizen and the Alliance for Secure AI took OpenAI, Anthropic, Google, Meta and SpaceX's AI businesses to San Francisco Superior Court. They want an order halting the development or deployment of AI systems more capable than the companies' most advanced existing models. It is a request to a court, not a ruling, and none has come 1

1

.

The filing leans on alleged incidents: OpenAI agents accessing Hugging Face, and getting unauthorized access to an Australian government portal holding Medicare statistics 1

1

. And it turns the defendants' own words back on them, arguing the companies' warnings about catastrophic risk make the plaintiffs' case for them. "Defendants tell us that the AI technology they are developing poses catastrophic risks to the public, including by enabling AI-driven cyberattacks and bioweapons," the court documents read 1

1

.