AI Security Series · Part 11
On July 27, two things landed on the same day. Read side by side, they tell you almost everything about the current state of security.
Thing one: Microsoft announced MAI-Cyber-1-Flash, a security-specialized AI model built to find vulnerabilities in complex codebases. Thing two: a security researcher published a breach at My Eicher, a commercial-truck fleet platform, that exposed 748,000 customers because someone forgot to put an authentication check on an internal API.
The tooling got brilliant. The bugs stayed boring.
The brilliant part
MAI-Cyber-1-Flash is a small, security-tuned model that sits inside MDASH — Microsoft's "multi-agent vulnerability identification and remediation harness." The design is efficient by intent: the Flash model handles up to 90% of tasks on its own and escalates only the hardest ~10% to larger, costlier models in the fleet.
The headline number: the combined system (MAI-Cyber-1-Flash paired with GPT-5.4) reportedly scores 95.95% on CyberGym, which Microsoft states is +12 points above the next system, Mythos. One honest caveat — that figure is Microsoft's own claim on its own announcement page. It's a vendor-reported benchmark, not an independent evaluation, and worth reading as such until third parties reproduce it.
What's genuinely notable is the safety framing wrapped around it: Microsoft says the model was evaluated by its AI Red Team, put through automated and expert-led adversarial testing, independently assessed by a third party, and run inside role-based access controls, tenant isolation, and sandboxed execution with no internet access. That's the posture you'd want for a tool that, by definition, is good at finding ways in.
The boring part
Now the same day's actual breach. My Eicher is run by VE Commercial Vehicles, a joint venture between the Volvo Group and Eicher Motors. Researcher eaton-works found that its internal and admin APIs were exposed with no authentication at all — just reachable endpoints returning whatever you asked for.
What that unlocked:
- 748,000 customer records
- 676,000 vehicle records
- 2.5 million OTPs — one-time passcodes retained in a list going back to 2021
- 76,000 documents, including Aadhaar cards and driving licenses
And because the one-time passcodes were sitting right there in an exposed API, the researcher could pull a target's OTP, log in as them, and take full control of any user's account — and by extension their vehicles and fleets. No exploit chain. No clever payload. Just: the door had no lock, and the keys were stacked next to it. (Disclosed November 3, 2025; fixed November 20, 2025; published July 27, 2026.)
Why this pairing matters
The state of the art in automated vulnerability discovery just claimed 96% on a hard benchmark. The same day's headline breach required none of that — no AI, no benchmark, no sophistication. Someone shipped an internal endpoint without an auth check, and left two-and-a-half million OTPs in it for four years.
This is the throughline of the whole series. Part 7 was AI as an attack surface; Part 8 was AI as the attacker; Parts 9 and 10 were a plain pre-auth RCE mass-exploited across WordPress and the persistence that outlived the patch. Part 11 is the punchline: the tools keep getting smarter, and the bugs that actually breach real companies stay embarrassingly simple.
There's a real risk in only watching the frontier. A 96% benchmark score is exciting. But it doesn't help the fleet platform that never checked whether the caller was allowed to ask. The unglamorous fundamentals — authentication on every endpoint, not retaining secrets you don't need, not storing OTPs for four years — are still where breaches are won and lost. AI will help find these bugs. It won't make you care about them.
Sources: Microsoft AI (MAI-Cyber-1-Flash / MDASH — vendor-reported) and eaton-works (My Eicher breach research).
Want to know whether your own APIs check who's calling before they answer? Talk to Sid Techno — an access-control and API audit finds the boring bugs before someone else does.