Frontier AI finds software flaws faster than the world can patch them
A week of outages at Microsoft, United Airlines, JPMorgan Chase and the NSA's classified network drew no cyber-attack blame, and the reason that should worry every board is the widening gap between how fast flaws are found and how fast they get fixed.
Garrison Lovely argues in The Guardian that frontier AI models can exploit software flaws faster than the world can fix them, a claim set against a week of outages at Microsoft, United Airlines and JPMorgan Chase, a nationwide grounding of US domestic departures and the reported failure of the NSA's classified internal communications network. For decision-makers, outage risk, disclosure risk and patch velocity now belong on the board agenda together, because none of the week's failures was publicly attributed to an attack.
Frontier AI models can exploit software flaws faster than the world can fix them. That is the claim at the heart of Garrison Lovely's argument in The Guardian, and the past week supplied the backdrop: outages took down core services at Microsoft, United Airlines and JPMorgan Chase. On Friday, all US domestic departures were grounded, the second time that has happened since 9/11. On Tuesday, CNN reported that the National Security Agency's highly classified internal communications network went down. In every case, officials or the companies said there was no evidence of a cyber-attack.
That last line is the one to read twice. "No evidence" describes what investigators have found so far, not a verdict on what happened. Cyber attribution historically takes weeks, sometimes months, and typically requires correlating logs, code and infrastructure across multiple jurisdictions. The uncomfortable implication is that when the best-resourced organizations on earth cannot immediately explain why their systems fell over, everyone downstream is flying blind too. Lovely's argument is not that a specific attack caused this particular week of chaos. It is that the tooling now exists to make the next one faster, wider and far harder to trace.
The core of the case is velocity. Software vulnerabilities have always been a race between the people who find them and the people who patch them, but historically that race ran at human speed: a researcher reports a flaw, a vendor triages it, a fix ships, IT departments schedule the update. Frontier models collapse that timeline on the offensive side, scanning code and drafting exploits at machine speed, while remediation still runs on change-management calendars, testing cycles and narrow maintenance windows. The gap between how fast a flaw can be weaponized and how fast it can be closed is the whole ballgame, and the source argues that gap is widening rather than narrowing.
Lovely also flags that the risks extend well beyond the cyber realm, and the week's casualty list shows why. Airlines, banks and the federal government's most sensitive communications all sit on the same basic plumbing: cloud infrastructure, identity systems, network providers, third-party software. When that plumbing stutters, the damage does not stay contained to one company or one sector. A grounded fleet becomes a scheduling, crew and revenue problem within hours. A bank outage becomes a trust problem. A classified network going dark becomes a national security problem. None of those outcomes requires a malicious actor to be consequential; they only require a system that nobody fully understands end to end.
For context, the disclosure environment has tightened considerably. The SEC adopted rules in 2023 requiring public companies to report material cybersecurity incidents within four business days of determining they are material, alongside annual disclosure of how they manage cyber risk. European regulators have layered on their own reporting obligations. That creates a specific tension for the companies in this story: saying "no evidence of a cyber-attack" quickly is both a legal and a reputational shield, but it is also a statement that can be revised. Boards should assume that whatever is said in the first 48 hours of an outage will be compared, line by line, with what surfaces weeks later.
The deeper structural issue is concentration. A handful of hyperscalers, a handful of identity providers and a handful of security vendors sit beneath thousands of companies. The July 2024 CrowdStrike update, which disabled Windows machines across airlines, hospitals and banks, remains the cleanest recent illustration of how a single upstream failure travels: one bad push, global blast radius, days of cleanup. Layer AI-accelerated vulnerability discovery on top of that architecture and the arithmetic gets worse. Defenders can reach for the same class of models, but they have to be right everywhere, every time. An attacker, or simply a cascading failure, only has to be right once.
So what should an operator or a board actually do with this? Treat patch velocity as a board-level metric rather than an IT afterthought. Ask how long it takes from a publicly disclosed vulnerability to a deployed fix, and how many critical systems are quietly exempt from that clock. Stress-test the assumption that "no evidence of an attack" is the same as "no attack." Insist that incident response plans account for failures with no adversary at all, because a week like this one suggests the unglamorous version is the likeliest. And expect the disclosure rulebook to keep tightening, because regulators historically move after the outages, not before them. The internet does not have to break all at once. It can simply get steadily harder to keep standing.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.



