There are glaring issues in this paper beyond the evaluation problems raised by Davi and ToB, including ones that make me question the ratio of human to AI assistance here, but a few are so egregious we should consider what research norms we are demanding from industry labs.

There's no reason you can't use an LLM to help you write. There's only one rule: you can't use a single word the LLM selects for you, even if it's better than your own; users detect LLM words in the parts per trillion. Here's what I've been using:

Corporate Infosec sends a phishing test email, I click, and somehow I'm the asshole because "[I] failed the test; had this been real it would have destroyed the company network"? If me clicking can destroy the network, I'm not the one in this conversation who sucks at their job.

.@1Password's FLAWED report says AI models produce a clean security fix only 26% of the time. Defenders shouldn't take that number seriously. • The six vulnerabilities were handpicked because their fixes were complex. Clean-fix rates ran from 3% to 60% depending on the bug, and the report averaged them together. • Agents set up to fail were counted in the headline figure. Two of 1Password's prompts instructed the agent to apply the wrong fix. Those trials make up 22% of the data. One evaluation mode prevented the agent from compiling or running any code, and it accounts for 36% of the data. • The report ran two models, GPT-5.5 at medium effort and Opus 4.8 at high. Neither was tested at its highest available setting, so the report says nothing about how more effort or stronger models change the results. • Several instruction and grading errors further undercut the headline, and are elaborated upon in the attached blog. We've spent four months submitting hundreds of AI-authored patches to widely adopted open-source projects as part of Patch the Planet. Our experience didn't match 1Password's report, so we did a full analysis across 186 AI-authored pull requests and 33,500 subsequent commits, benchmarked against 2,265 human-authored patches we graded across years of security engagements. blog.trailofbits.com/2026/09…

Improved key recovery attacks against Classic McEliece by @sweis with help from Claude All standardized parameter sets for Classic McEliece are no better than NIST Level 1 now eprint.iacr.org/2026/1984

.@1Password's FLAWED report says AI models produce a clean security fix only 26% of the time. Defenders shouldn't take that number seriously. • The six vulnerabilities were handpicked because their fixes were complex. Clean-fix rates ran from 3% to 60% depending on the bug, and the report averaged them together. • Agents set up to fail were counted in the headline figure. Two of 1Password's prompts instructed the agent to apply the wrong fix. Those trials make up 22% of the data. One evaluation mode prevented the agent from compiling or running any code, and it accounts for 36% of the data. • The report ran two models, GPT-5.5 at medium effort and Opus 4.8 at high. Neither was tested at its highest available setting, so the report says nothing about how more effort or stronger models change the results. • Several instruction and grading errors further undercut the headline, and are elaborated upon in the attached blog. We've spent four months submitting hundreds of AI-authored patches to widely adopted open-source projects as part of Patch the Planet. Our experience didn't match 1Password's report, so we did a full analysis across 186 AI-authored pull requests and 33,500 subsequent commits, benchmarked against 2,265 human-authored patches we graded across years of security engagements. blog.trailofbits.com/2026/09…