Can AI find security flaws in software?
YesYes. AI systems have found serious, previously unknown security flaws in widely used software, and in October 2026 AI-found flaws in PHP led governments to switch services off as a precaution.
First shown: Nov 1, 2024. Google's Big Sleep agent found a previously unknown, exploitable flaw in the SQLite database engine, fixed before any release was affected; Google called it the first public example of its kind. (Google Project Zero, Nov 2024)
What it means
Finding security flaws used to take scarce human experts weeks. AI models can now read large amounts of code and spot mistakes that automated testing missed, and labs report finding thousands of serious flaws this way.
The same ability cuts both ways. Defenders use it to find and fix bugs first; that is how most of these flaws were found. But once a flaw is public, attackers can use AI to turn it into a working attack faster, which is part of why the PHP flaws were attacked within days.
The most capable bug-finding tools are mostly restricted to vetted security teams. The general models anyone can use are less capable at this, but improving.
The evidence ladder
Every capability on this site is graded on the same five steps, from a single demonstration to real-world consequences.
| Step | Status | When | Evidence |
|---|---|---|---|
| Demonstrated shown at least once | yes | Nov 1, 2024 | Google called it the 'first public example' of an AI agent finding an unknown exploitable memory-safety issue in widely used real-world software, and described the work as 'highly experimental'. (Google Project Zero) |
| Reproduced independently repeated | yes | Apr 7, 2026 | Anthropic's Mythos Preview reported thousands of high- and critical-severity vulnerabilities, including in every major operating system and browser (Apr 2026); OpenAI's Daybreak program said its cyber-tuned model found two high-severity Chrome bugs (Aug 2026). (Anthropic) |
| Accessible available to ordinary people, or criminals | partly | Oct 6, 2026 | Frontier bug-finding is gated to vetted defenders: Anthropic's Glasswing and three-tier Cyber Verification Program, OpenAI's Daybreak tiers with hardware keys from Sept 2026; general models such as Opus 4.6 are still said to 'find real vulnerabilities in codebases'. (Anthropic) |
| Reliable works most of the time for a defined task | partly | Oct 23, 2026 | Forbes, Oct 23: OpenAI's Patch the Planet program (run with Trail of Bits) has found 1,995 bugs, 1,316 of them still awaiting a patch; of 130 AI-written patches deployed, 16 caused functional problems and one a performance problem, none a new weakness. Anthropic's Glasswing has found more than 29,000 possible bugs and triaged about 6,000. No published success rate for discovery of flaws of PHP2Shell's severity. (Forbes, Oct 23, 2026) |
| Consequences real-world use or harm documented | yes | Oct 26, 2026 | PHP2Shell: governments from the Netherlands to Japan took public services offline as a precaution; millions of exploit attempts. Unit 42's Oct 23 report of a successful intrusion at a European government body was retracted on Oct 25 (it could not tie the break-in to these flaws); as of Oct 26 code execution is confirmed only on researchers' decoy servers and no organization has confirmed a breach. (The Hacker News, Oct 2026) |
From the record
- Google's Big Sleep finds a previously unknown flaw in SQLite, fixed the same day. (Google Project Zero)
- Google reports a criminal group using an attack on an unknown flaw that it assessed was likely developed with AI. The case file → (TechWire Asia (Google GTIG))
- Researchers use frontier AI to find critical flaws in PHP, the software under most websites; governments take services offline as a precaution. The case file → (Wiz Research)
What it means for you
The software you rely on is being checked for flaws faster than ever, by both defenders and attackers. The practical consequence for you is that updates matter more and arrive more often.
What to do
- Turn on automatic updateson your phone, computer, browser and router. Fixes for newly found flaws now arrive fast; installing them quickly is the best protection.
- If you run a website,make sure whoever hosts it applies security updates promptly, and ask which software versions it runs.
- Use two-step loginon important accounts, so a flaw in one service doesn't hand over everything.
Questions people ask
Can AI find zero-day vulnerabilities?
Yes. A zero-day is a flaw nobody knew about. Google's Big Sleep found one in 2024, and labs have since reported many more, including the PHP flaws disclosed in October 2026.
Is AI making hacking easier?
It makes some steps faster for attackers, especially turning known flaws into attacks. It also helps defenders find and fix flaws first, which is how most AI-found flaws have been handled so far.
What should I do to stay safe?
Keep everything updated automatically, use two-step login, and be wary of messages that use security news as a lure.
Related
Sources
- Google Project Zero: From Naptime to Big Sleep (Nov 2024)
- Anthropic: Mythos Preview
- Wiz: PHP2Shell
- Forbes: OpenAI's plan to protect the internet (Oct 23, 2026)
- The Hacker News: PHP2Shell exploit attempts (Oct 2026)
Spot an error or something out of date? Tell us; corrections are made and logged. Written with AI assistance from public sources; every fact checked and approved by Colin Balfe. General information, not legal, medical or financial advice.