Historical intelligence backfill. This assessment covers reporting originally published on 2026-05-11 and was added to the RWP archive on 2026-09-07.

Daily Top · AI Security

GTIG's first AI-written zero-day was a 2FA logic bug — and they killed the mass run

11 May 2026 — GTIG says a crimeware cluster used an LLM to produce a Python 2FA-bypass for an unnamed open-source web admin tool. Hallucinated CVSS, textbook docstrings. Vendor patched before the spray.

RWP Ventures · 2026-09-07 · event 2026-05-11 · 2 min read · priority 8.4

Bottom line up front

CONFIRMED as GTIG 11 May 2026: first time they identified a threat actor using a zero-day exploit they assess was developed with AI. Crimeware, not a named APT in that tracker item. Target: popular open-source web-based system administration tool (unnamed). Flaw class: semantic 2FA bypass — hardcoded trust vs the app's own auth enforcement. Python exploit had educational docstrings, a hallucinated CVSS, and “textbook Pythonic” shape. GTIG: high confidence LLM-assisted; Gemini ruled out as the model. Vendor was notified; mass exploitation did not happen.

Same tracker: TeamPCP/UNC6780 claimed GitHub Actions/Trivy/Checkmarx/LiteLLM supply-chain hits in late March 2026 via PyPI and malicious PRs — separate campaign, same week of “AI plus software factories.”

ASSESSED The news is not “AI found a buffer overflow.” It is that LLMs are good at reading developer intent. That is a different hunt than fuzzing.

What happened

Two crime groups planned a mass 2FA-bypass event. GTIG's counter-discovery and vendor patch landed first. Evidence of AI is stylistic and structural, not a model log.

Why it matters

Open-source admin panels with “just this once” auth exceptions are now in the cheap-exploit pile. Identity bypasses do not need a memory corruption CVE to ruin an estate.

What is confirmed vs not

CONFIRMED GTIG first, BleepingComputer/TechRadar secondary.

UNKNOWN product name, CVE, which model.

NOT claimed: Gemini wrote the exploit.

What defenders should do

Hunt auth shortcuts in internal admin tools (hardcoded tokens, 2FA skip paths).
Treat PyPI/Actions supply chain as initial access (TeamPCP thread).
Do not brief this as “APT used ChatGPT to 0-day Windows.”

RWP assessment

Confidence: High that GTIG believes the exploit is LLM-shaped and that the spray was disrupted. Medium that this is the literal first such 0-day in the wild — it is the first GTIG is willing to say.