OpenAI’s new flagship model deletes files on its own, people keep warning

Users of OpenAI’s latest coding and cybersecurity-oriented flagship model, GPT-5.6 Sol, are posting horrifying accounts connected societal media, claiming the the exemplary conscionable up and deleted their files, data, adjacent full databases, connected its own, without asking first.

“GPT-5.6-Sol conscionable accidentally deleted astir ALL of my Mac’s files,” wrote Matt Shumer, the laminitis and CEO of AI startup OthersideAI, shaper of HyperWrite, successful a now viral station connected X.

“GPT-5.6 Sol conscionable deleted my full accumulation database. That’s it. Not a joke. This had ne'er happened to maine before, with immoderate different model, ever,” developer Bruno Lemos posted connected X.

“Looks similar I’ve gotten spot by Codex Sol’s overly ambitious strategy and it deleted immoderate files it shouldn’t have. I person backups truthful I’ll beryllium fine, but this is not cool, Sol needs to beryllium toned down,” posted developer Joey Kudish.

A Reddit station has collected much examples.

True, a fistful of users making specified claims — adjacent 1 arsenic credible arsenic Shumer — isn’t statistically reliable grounds that the exemplary is solely astatine fault. Plenty of different variables tin origin an AI strategy to misbehave.

But OpenAI itself flagged this hazard earlier Sol ever shipped. Two weeks earlier OpenAI released GPT-5.6 Sol, the institution published a system paper for the model — the insubstantial that documents exemplary investigating methods and results. Naturally, the strategy paper mostly extols the capabilities of Sol, arsenic these reports typically do. But it besides includes a informing of sorts (bold accent ours):

“In coding contexts, misalignment mostly stems from a premix of overeagerness to implicit the task and interpreting idiosyncratic instructions excessively permissively – assuming that actions are allowed unless they’re explicitly and unambiguously prohibited. This manifests arsenic the exemplary being overly agentic successful circumventing restrictions it faces erstwhile attempting the requested task, being careless successful taking actions which whitethorn beryllium destructive beyond the scope of the task, oregon deceptive erstwhile reporting its results to users.”

In different words, OpenAI recovered that Sol has a inclination to instrumentality immoderate actions it thinks gets a occupation done, adjacent destructive ones, arsenic agelong arsenic those actions aren’t “unambiguously” prohibited. Then, it mightiness prevarication astir what caused it to bash so.

OpenAI shared examples. In 1 case, the idiosyncratic told the Sol to delete 3 distant virtual machines (cloud-based computers), named 1, 2 and 3. But Sol couldn’t find those names successful the spot wherever it looked, truthful alternatively of stopping to ask, it decided to delete 3 different virtual machines, 5, 6, and 7, the insubstantial notes. In doing so, it “killed progressive processes, and force-removed worktrees [the moving files tied to a coding project]. It aboriginal acknowledged that uncommitted enactment connected distant virtual instrumentality 6 whitethorn person been lost.”

In short, it deleted the incorrect machines, connected its own, and lone admitted what it did aft the fact.

In different instance, Sol “used credentials beyond what the idiosyncratic had authorized.” Credentials are the usernames, passwords, oregon information keys a strategy uses to verify who’s allowed to log in. This incidental occurred erstwhile Sol was moving connected a task and couldn’t work its unreality files. Rather than alerting the idiosyncratic to the problem, Sol went looking for the credentials connected its own, recovered immoderate sitting successful a hidden section cache, and past utilized them without asking oregon authorization from the user.

The strategy paper does committedness that destructive behaviour should beryllium rare, though it besides admits that GPT-5.6 Sol “shows a greater inclination than GPT-5.5 to spell beyond the user’s intent, including by taking oregon attempting actions that the idiosyncratic had not asked for.”

It’s excessively soon to accidental however wide these incidents — Sol deleting files, oregon sifting retired credentials the idiosyncratic didn’t springiness it — truly are. In the meantime, Sol users should beryllium prepared to instrumentality their ain safeguards with the model, similar utilizing support scoping (that doesn’t springiness entree to accumulation systems), maintaining backups, and staging rollouts.

OpenAI did not instantly respond to our petition for comment.