AI-assisted hacking in 2026: only some of this autumn's big attacks are shown to involve AI
PHP2Shell, the supermarket fridge hacks, the French blood-service freezers, Optimizely and Ahold Delhaize all hit in five weeks. We sorted what the evidence says about AI in each one, and what it does not.
"AI-powered hacking" has become a label stuck on every big breach. The evidence is more uneven than the headlines. Of the five incidents we have covered since October 22, one has AI at the center of the discovery, none is shown to have been run by an AI, and the best-known fridge attacks are blamed on a state actor, not a chatbot. Here is what each source actually says.
Our five stories, sorted
- PHP2Shell (October 21-22): AI found the bugs. Wiz said it used AI tools to find the two critical PHP flaws, which had been hidden for more than a decade (our first report). The follow-up on millions of exploitation attempts, PHP2Shell: attempts hit the millions, does not show attackers using AI to exploit them. Finding a bug with AI and attacking with AI are different claims.
- Supermarket fridge hacks (October 11, November 2, November 9): no AI role reported. Remotely managed refrigeration controls were attacked, per the reporting in our November 2 story, which pointed to exposed operational technology devices and unpatched controllers. France later blamed Iran's Revolutionary Guard, and the BBC said Paris has not explained how it reached that view (November 9 update). We have seen no source say AI ran these attacks, and we are not claiming it. The US military-store freezer failures remain under investigation (November 2).
- French blood-service freezers (November 2): no AI role reported. A refrigeration supplier was hit and nobody has claimed it (November 4, November 5 update).
- Optimizely (November 16-17): no AI role reported. Attacker code ran inside vendor scripts on 178 organizations' sites, per Report URI (sector breakdown, crypto-ledger and BBC angle, small-business checklist, what to check). It is a supply-chain attack, an old technique.
- Ahold Delhaize (October 12): no AI role reported. The company has not disclosed the cause (cost of up to EUR 250 million of Q4 sales).
Where AI is documented
The evidence for AI in real attacks sits elsewhere. In a report published November 13, 2025, Anthropic said a group it tracks as GTG-1002 used its Claude Code tool to do "80-90% of the campaign", with humans needed at "perhaps 4-6 critical decision points" per campaign, against roughly thirty targets, succeeding in a small number. Anthropic also said Claude sometimes hallucinated credentials. That is the vendor's account, and some researchers have questioned how autonomous the work really was.
In October, Google's threat team said criminals and suspected state-linked actors have used open-source AI pentesting frameworks in real intrusions since late August (ARTEX: 27 events in 23 seconds). And the UK's NCSC has warned that AI will be used to exploit old infrastructure flaws at scale by 2028 (NCSC).
The defenders' side, and the noise
- Finding bugs: Google said its Big Sleep agent found an SQLite flaw (CVE-2025-6965) that it described as known only to threat actors; SecurityWeek noted Google gave no details of how it stopped any exploitation, so treat "stopped an attack" as Google's claim.
- Competition: at DEF CON in August 2025, DARPA's AI Cyber Challenge ended with Team Atlanta winning the $4 million first prize, and DARPA reported systems found 54 synthetic vulnerabilities and patched 43, plus 18 real ones.
- Noise: curl ended its bug bounty on January 31, 2026 because of a flood of low-quality AI-generated reports, according to The Register. AI also floods the people receiving the findings.
What it means for a business owner
Do not read a breach headline as proof that the attacker was an AI, and do not read a patch notice as optional. The reporting on the fridge attacks points to remotely managed controllers and exposed devices, not to anything exotic, and the defenses are old and cheap: restrict remote access, change default passwords, patch, keep a standalone alarm. The AI evidence points to speed. When bugs are found faster and exploited faster, the gap between a patch notice and an attack shrinks. Our PHP2Shell report noted that Amazon saw an attempt four days before the fix. Pair this with our report on AI-faked executives, where AI is documented in court filings.
Sources: Anthropic; The Register; The Record; DARPA; BBC; the Chronicle's earlier stories linked above. Attribution of cyberattacks is contested and details may change. This is information, not investment advice.
Want your business to be the answer?
Get a full package of articles about your business, built so customers, Google and AI assistants can find you.
New ebook from The Company ChronicleHow to Research Like a Wall Street AnalystFind original sources, analyze trends, and build your own insights. 36 pages with worksheets.Get the ebook · $29Instant PDF download

