The case
On October 21, emergency patches shipped for two vulnerabilities in PHP itself, the language running behind a good part of the web.
They were registered as CVE-2026-109618 and CVE-2026-109619 with a CVSS of 9.2, and the pair got the name php2shell (Source: fortune.com).
Both are remote code execution, meaning flaws that let an attacker run whatever they want on the server without authenticating first.
What stands out is not the flaw itself but where it came from.
The code had been sitting there for over a decade, invisible, and it was found with help from AI (Source: fortune.com).
Cloudflare already confirms active exploitation.
And several UK and European government sites, plus Wikipedia and the Internet Archive, went down preventively (Source: fortune.com).
That preventive shutdown is the part worth reading twice, because it was not an attack that took those services down but the decision to switch off before it arrived.
Not an isolated case
The same week brought another flaw with the same profile.
Oracle E-Business Suite has an unauthenticated SQL injection in the TCF servlet, registered as CVE-2026-101767 with a 9.8 that chains up to remote code execution (Source: helpnetsecurity.com).
It was patched in the October 20 CPU and there are already exploitation probes.
Censys counts 2,246 web properties with EBS exposed (Source: helpnetsecurity.com).
Old flaws, in software nobody looks at because it has been running for years, found more easily now and exploited within days.
In parallel, CISA added PolyShell (CVE-2026-48356, 9.3) and Node2Shell (CVE-2026-111497, 9.2) to KEV. Sansec sees more than 950 Adobe Commerce and Magento stores with a loader in CMS blocks since July 2025, and there are already Brotli exploitation attempts against honeypots (Source: thehackernews.com).
What used to take months to go from flaw to exploit now takes hours.
The other side
AI does not only find flaws.
Moonshot removed cyberattack material from training before opening Kimi K3.1 Pro, and the model still solves 81% of its internal exploit tasks, against 33% for the previous version (Source: thenextweb.com).
Offensive capability does not travel in the dataset, I think it travels in the access.
Filtering data is no substitute for limiting permissions or monitoring what the agent does with them.
Lazarus, the North Korean group, encrypts the Operation FlashHole malware with post-quantum cryptography, active since July 2025 against South Korean construction, software and cable manufacturing companies (Source: thehackernews.com).
The vector is a watering hole, meaning compromising a site the victims visit regularly, in this case the mandatory security software iniLINE Cross EX (Source: thehackernews.com).
My reading is that the three pieces tell the same story from different angles:
- The attacker speeds up in three phases: finding, weaponizing and hiding.
- The defender still measures speed in the fourth one: patching.
My own reading
For years the debate was whether AI would find flaws humans could not see.
The answer has arrived and it is yes.
Finding a decade-old flaw in hours is real progress, but if the patch takes weeks to reach production the progress only benefits whoever exploits first.
It comes down to the inventory of what runs in production, and update speed in most organizations is measured in months, weeks if you are optimistic.
What to look at
- The real inventory. Not the CMDB one, the one that answers a scan. If you do not know which PHP version runs on each host, the October 21 patch is useless to you.
- The patching window. How long a critical patch takes from release to production. If the answer is more than a week, the decade-old flaw has already reached you.
- EBS exposure. If you have Oracle E-Business Suite with the TCF servlet reachable from the internet, that is the point to close before thinking about the rest.
How would I test it in my lab?
I would spin up a container with an old PHP version, one of the ones carrying the flaw, and leave it on an isolated network.
I would run a scan with a model-assisted static analysis tool and measure how long it takes to flag the vulnerable function, and I would bet you do not need frontier models.
Then I would time how long it takes me to apply the patch and bring the service back up.
The question here is how long I take to close it.
Closing
AI has made finding flaws cheap but not fixing them, and while discovery is measured in hours and patching in weeks or months, the advantage always goes to whoever attacks.
Originally published at https://sammideblas.com/notas/ai-made-finding-bugs-cheap-not-fixing-them
Si has leído hasta aquí... reacciona y comparte -> "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?
Top comments (0)