The Super Intelligence Task Force should open the door to defenders
Artificial intelligence models now find software vulnerabilities faster than organizations can fix them. In HackerOne’s platform data, the valid report backlog is up 131% over the last two years, and critical and high-severity findings are rising while remediation falls further behind. Finding flaws is no longer the hard part. The harder question will decide the balance between offense and defense in the age of superintelligence: When someone outside an organization finds a flaw in a system Americans depend on, is anyone on the other end listening?
The Task Force on Super Intelligence is about to give that question an official answer, whether it means to or not. Its charter directs it to commission a report on the impacts of superintelligence on “the offense-defense balance in cybersecurity.” It also requires the task force to assess the preparedness of the government and the private sector across healthcare, communications, banking, energy, water, transportation and emergency services. Both assessments will shape federal policy for years. If they leave out third-party testing and disclosure, they will start from an incomplete picture.
Recognizing third-party evaluation and disclosure
Any honest assessment of the offense-defense balance must ask how quickly a defender learns about its own weaknesses. Internal testing is necessary, but it has limits. The people who build and run a system will often share its assumptions, and they may miss what an outsider sees right away.
That is a central finding of a paper on third-party AI flaw disclosure I co-authored with researchers from Stanford, Princeton, MIT and elsewhere. Experts in software security, machine learning, law and policy agree that independent researchers need a clear way to report flaws, and they need legal protection when they do.
The task force should carry that insight into its work. Whether an organization accepts, triages and acts on reports from outside researchers, such as through a vulnerability disclosure program, is one of the best available indicators of defensive maturity. It is not perfect, but it shows something no internal audit can: whether an organization has accepted outside scrutiny, has a working process for handling what it hears, and turns findings into fixes. Frontier models will give attackers enormous scale. Defenders need scale of their own, and the global security research community, now working alongside AI, is the largest pool of defensive talent available. An assessment that measures attacker capabilities but not defenders’ ability to receive outside assistance is looking at half the board.
A foundation, not an extra
The federal government has already settled the threshold question. The Office of Management and Budget has required agencies to accept outside vulnerability reports since 2020. This June, an executive order moved to extend that expectation to covered federal contractors, recognizing that federal networks are only as resilient as the companies that support them. Congress is moving the same way through the bipartisan Federal Contractor Cybersecurity Vulnerability Reduction Act, included in the House version of the fiscal year 2027 National Defense Authorization Act. The direction is clear. Accepting outside reports is no longer an optional best practice. It is operational infrastructure.
The task force’s preparedness assessment should apply that logic to every sector named in its charter. Hospitals, banks, energy and water utilities, communications networks and emergency services are where one unreported flaw can become a national emergency. As frontier AI shortens the time between discovery and exploitation, the cost of having no way to hear a warning only grows. A rural hospital, a community bank or a local water utility that cannot receive a report from a good-faith researcher is not prepared, whatever else it has in place. A framework that inventories sectors without asking whether each operator can hear a warning gives a false sense of readiness. That is worse than incomplete. It is dangerous.
What the task force should do
The charter gives the task force 120 days and a mandate to coordinate across government and industry.
Within that time, it should make outside reporting and testing an explicit benchmark in the offense-defense report, measuring how quickly reports are triaged and fixed, not just whether a channel exists. It should also map where independent reporting is and is not available across every critical infrastructure sector in its charter, and treat the gaps as preparedness findings.
It should extend the same approach to AI systems themselves, with standardized flaw reports and a clear path for researchers to report model flaws and unexpected outputs. And it should recommend clear legal safe harbors, so the researchers this system depends on are not exposed to legal risk for doing the right thing.
The Pentagon showed a decade ago that even the most sensitive institution in the country could be safer by inviting outside experts in. The lesson matters more now. The task force charter says emerging risks should be “identified and communicated to government personnel, industry, and the American people.” Outside researchers are how that identification happens in practice. The task force should build outside researchers’ work into both assessments from the start.
Originally published on CyberScoop