AI Made Old Security Mistakes Cheap to Find
7 min read
#security #ai #open-source #disclosure #career
A password in an app
On 30 Oct 2026, TechCrunch reported that Applebee's Android app contained a password for a customer database. Researcher Robert Wiggins used AI to help find it, and TechCrunch confirmed it by decompiling the app. The flaw is fixed. Parent company Dine Brands said it had found no evidence of unauthorised access so far.
Anthropic's September threat report describes the same hunt at scale. One operator it groups under suspected ShinyHunters affiliates downloaded 1.8 million Android APKs, decompiled them and scanned them for hardcoded secrets with TruffleHog. The report does not say a model ran that step.
I think AI did not invent new attacks. It made old security mistakes cheap to find, and the hard part has moved to fixing them. Security stays a growth field, but demand is moving towards people who can fix and judge, and the way in is getting narrower.
The attacks are familiar
Anthropic's summary: "None of the operations in this report depended on some entirely novel technique that defenders have never seen. Instead, the economics of the attacks have changed." Also: "Several of the most serious compromises we report here came from operations where a human directed every step."
The 2026 Verizon DBIR agrees from different data: most AI-assisted malware and tooling development "was associated with well-known and defined attack techniques".
WorkOS shows the pattern in one incident. Its 06 Nov 2026 update says a researcher's AI agent escaped a test environment and reached an internal tool holding sales and support data. That is WorkOS's account, validated by an independent forensics firm. The way in was an internet-facing API, a predictable credential meant for local development, and no rate limiting.
Umut Bayram of Picus writes that CVE-2026-119977 had been in the core updater since WordPress 2.7 in 2008, and that he went from first look to working exploit chain in about four hours with GPT 5.6 Sol. WordPress 7.1.4 fixed it on 28 Oct 2026. Xint says it had confirmed the same flaw independently on 22 Oct 2026.
The queue moved to the maintainers
In January, Daniel Stenberg ended the bug bounty after confirmed reports fell below 5% amid AI slop. By October the complaint had changed: "Since spring 2026 the slop reports are practically gone. Our current challenge is much harder: high volume, high quality." Two days earlier curl had announced that LOW and MEDIUM reports would go to public issues from 26 Oct 2026, saying key maintainers were "brought to the brink of what we can muster".
Linus Torvalds wrote on 17 May 2026 that the kernel security list had become "almost entirely unmanageable, with enormous duplication due to different people finding the same things with the same tools". His summary: "if you found a bug using AI tools, the chances are somebody else found it too."
I reported the OAuth credential issue behind CVE-2026-104850 in the MCP TypeScript SDK on 17 Apr 2026. Six other people reported it separately, and the advisory credits all seven. I don't know how the others found it, so this doesn't prove Torvalds's point. From the reporter's side, duplication looks like seven reports to read and one fix to get right. That fix shipped on 28 Sep 2026.
Numbers that disagree
Anthropic's Glasswing programme has the largest count I have seen, and it is contested. On 22 May 2026 Anthropic said 75 of the 530 high or critical bugs it had reported were patched; its dashboard showed 97 of 1,596 across all severities. It said it was early in its 90-day window and probably undercounting, but that the low volume "reflects a genuine problem". Its page, as of 02 Nov 2026, says 5,087 disclosed and 629 patched, about 12%. That isn't a clean fix rate: 3,225 of the 5,087 went to maintainers without external review.
Patrick Garrity at VulnCheck disputed the numbers on 08 Sep 2026. The dashboard said 421 patched and the ledger said 202: "they just don't reconcile". Where both had rated a finding, Claude called 91.5% critical or high and maintainers 51.3%. The gap persists. The 02 Nov 2026 data file has total_fixed_in_response at 629 and total_fixed at 217. Patch flags in the ledger appear only on findings with published details, so my reading is that outsiders can't yet check the rest. Anthropic hasn't said that. On either count, fixes trail reports.
Neither of two older figures shows AI caused the gap. The DBIR's median time to fully patch known-exploited vulnerabilities rose from 32 to 43 days, on data running to November 2025. M-Trends 2026 estimates mean time to exploit at -7 days, and adds "we do not consider 2025 to be the year where breaches were the direct result of AI". Fixing was slow before finding got cheap.
Who said this first
This argument isn't mine. Thomas Ptacek wrote in March that agents would alter "both the practice and the economics of exploit development". The Cloud Security Alliance wrote in June that the bottleneck "has shifted from vulnerability discovery to vulnerability remediation". What I can add is narrow: attacker, maintainer and hiring evidence in one place, with the disagreements left in.
The case against
Most of these numbers come from companies that sell defence. Anthropic sells the models and Claude Security, runs Glasswing, and is cited throughout this post. Picus sells attack simulation and Xint sells AI-powered discovery. I weight curl and the kernel highest, because those are maintainers describing their own inboxes. The vendor figures I read as claims.
The stronger objection is that this tooling helps defenders first. Mozilla says Firefox 150 fixed 271 vulnerabilities from one Claude Mythos Preview evaluation, and wrote "Defenders finally have a chance to win, decisively." Stenberg credits Aisle, using AI tooling, with 35 published curl CVEs. I accept both. But Mozilla warned teams they may need to "reprioritize everything else", and Aisle's engineers wrote curl commits. The gain arrived where someone owned the fix. If models start writing reliably correct patches, the bottleneck moves again and this post ages badly.
The way in
Demand still grows. AI Workforce Consortium data, published by Cisco, has G7 cybersecurity job demand up 9.5% in the six months to March 2026. Senior-titled postings grew 65.0% and junior-titled 5.9%. The bases are small: 6.7% and 0.7% of postings.
ISC2 surveyed 856 professionals who use AI in May 2026. 56% said AI had reduced the need for entry-level positions, 53% said it is creating new types of entry-level role, and 62% said the need for foundational skills has not fallen. These are AI users' perceptions, not hiring data. ISC2 reads them as "not necessarily narrowing, but perhaps evolving". I'm less cheerful. The tasks it lists as moving to AI, such as alert triage and log analysis, "traditionally formed part of entry- and junior-level roles". Meanwhile 65% spend more time deciding when to trust AI recommendations. That judgement comes from having done the basics.
What to do
If you are starting out, stop counting bugs found. Reproduce a published advisory in your own lab, read the fix, then write the regression test that would have caught it. Check that it fails for the right reason.
Score a public report against the project's threat model before reading the maintainer's rating. Stenberg says AIs are "still bad at figuring out exactly where the line is between a bug and a vulnerability".
If you report bugs, assume someone else has it too. Search the public tracker, then send a short report, a tested reproducer and a patch.
If you run systems, list every internal tool the internet can reach.
If you ship an app, decompile your latest build this week and run a secret scanner over it. If it finds something, rotate the credential first. Removing it from the next release does nothing for installed builds.