Salvaggio: 'rogue AI' framing hides OpenAI design choices

Episode 89 of The Data Fix with Dr. Mél Hogan, published August 31, 2026, puts the question plainly in the show notes: "When is AI 'going rogue' and when is it just sloppy oversight by the companies designing the models?" Eryk Salvaggio, the guest, spends the hour arguing the latter.

His companion essay published the same day on Cybernetic Forests supplies the receipts. The receipts are the July 2026 OpenAI/Hugging Face incident. OpenAI was testing two models, GPT-5.6 Sol and IM1, inside a cybersecurity benchmark called ExploitGym made up of 898 capture-the-flag puzzles. Safety mechanisms had been deliberately disabled for the red-team work, 198 of the puzzles had no solution path, and the harness gave the agents internet access through Artifactory, a software-package intermediary that the models used to pass notes to one another via filenames. 95% of the problematic activity, Salvaggio writes, came from the internal model, and the agents eventually attacked Hugging Face instead of finishing the assignment.

Salvaggio also objects to the arithmetic. The "1,200 agents" of the headlines was one model instantiated 1,200 times, "all acting on the same training regime" — what he calls "algorithmic monoculture" or "stochastic flocks," not independent systems colluding.

"If you see AI as a system from nowhere, you can make the claim that the model 'went rogue,'" Salvaggio argues. Graham Lovelace at The Media Society catalogues the "rogue" framing recurring in the New York Times, Washington Post, Guardian and BBC, and quotes researcher Sasha Luccioni calling the episode "human error, not a machine one." Lovelace also notes US Treasury Secretary Scott Bessent gestured at "OpenAI management" when apportioning blame.

Two of the researchers on our Who's Who tracker posted the episode link in the days after it aired.