FLAWED cites 19 sources vs PatchBench's 73, exposing AI research gap

For the AI research community, FLAWED is a referendum on evaluation hygiene. The benchmark claimed "very little prior work" while citing 19 sources — mostly corporate blogs and an XKCD — against PatchBench's 73, a red flag that the authors did not survey the literature they dismissed. As frontier-model claims increasingly shape safety and capability narratives, the field must decide whether well-distributed industry papers get to skip the citation and review standards applied to everyone else.

The controversy over 1Password's FLAWED benchmark escalated on September 17, 2026, when a security researcher quote-tweeted Trail of Bits's post "1Password's AI patching benchmark is misleading," which itself pointed to Davi Ottenheimer's "Disinformation Pushed by 1Password: Their AI Patching Report is False." Both pieces challenged "Frontier Models' Vulnerability Patches are Often F.L.A.W.E.D" (FLAWED), a benchmark from 1Password's Off-by-1 Labs that claimed frontier AI models frequently produce flawed security patches. The suhacker.ai post published on September 24 moves the debate from "the benchmark is wrong" to a broader question: what research norms should the industry demand from well-resourced labs?

“

In the meantime, the safest posture for practitioners is to treat FLAWED as a case study in how not to evaluate frontier models, and to weight independent, citation-rich work such as PatchBench accordingly.

According to the author, who writes anonymously under a disclaimer that the views are their own, FLAWED contains "glaring issues" beyond the evaluation problems Trail of Bits and Ottenheimer raised — errors severe enough to raise questions about "the ratio of human to AI assistance" in the paper itself. The most serious problem is citation quality. FLAWED adopts the form and rhetoric of rigorous research, even claiming there is "very little prior work" in this specific area, yet it cites only 19 sources — mostly corporate blog posts plus an XKCD comic. The concurrent PatchBench paper, by contrast, cites 73 sources. That gap is not cosmetic; the "very little prior work" claim actively erases a body of literature the author argues clearly exists.

The author also documents arithmetic errors — one passage describes a value of roughly 4 when the actual figure is 2.8 — incorrect diagrams (including an asterisk in figure 1 that does not correspond to any step), and internal textual inconsistencies visible on a skim. Individually these might be dismissed as sloppiness; collectively, the author argues, they signal a paper that did not receive the scrutiny its distribution and branding implied.

The distribution problem is the heart of the critique. 1Password's marketing reach carried FLAWED into news coverage and, more consequentially, into defender roadmaps — the operational plans security teams use to decide where to invest. The author writes that watching this work "obscure more rigorous research from less-resourced groups" was what compelled the public post. A well-funded industry lab can out-shout independent academics whose more careful benchmarks never reach a CISO's desk.

The post also surfaces a quieter cost: after publishing, CISOs, academics, and industry researchers messaged the author privately to say they had noticed the same problems but stayed silent because they "lacked a forum or feared backlash." The author highlights second-order costs for academics in particular, who may depend on industry relationships for funding, data access, or career mobility, making public criticism professionally risky.

The author is explicit that the critique is not tribalism. "Research is not sports," the post argues — the fact that FLAWED was critical of OpenAI does not exempt it from scrutiny, and OpenAI is not the Spurs while 1Password is not the Knicks. The post also points forward: the same rigor should apply to claims inherent to Patch the Planet, and groups like EleutherAI demonstrate that precise evaluation of frontier-lab claims is achievable.