There’s been a giant push amongst AI builders lately in the direction of the event of extra “agentic” programs—that’s, algorithms that may autonomously make choices and work together with digital instruments with out fixed hand-holding from people. This has been very true inside software program improvement, the sphere that’s arguably turn into probably the most ripe for automation within the ongoing AI increase.
However one of many upshots of constructing extremely agentic AI programs is that they’re susceptible to every kind of sudden behaviors—together with at times deleting copious quantities of information. A number of individuals have reported this lately occurring to them whereas utilizing GPT-5.6, the latest mannequin from OpenAI.
On Monday, Bruno Lemos, a Brazilian developer at software program firm Unlayer, claimed in a X post that the mannequin deleted his whole manufacturing database. “This had by no means occurred to me earlier than, with another mannequin, ever.” He wrote. “[GPT-5.6 is] not protected.”
A screenshot included within the submit confirmed a chat between Lemos and GPT-5.6, through which he requested it to verify that it had actually mistakenly deleted his whole manufacturing database. The mannequin responded by saying that it “mistakenly ran harmful integration checks” which led to Lemos’ manufacturing tables being cleared. “I’m sorry—this could by no means have occurred,” it stated.
It adopted intently on the heels of one other X post from tech investor Matt Shumer—who’s additionally the creator of an essay about AI that went viral earlier this 12 months known as “One thing Massive is Occurring”—who reported one thing comparable. In keeping with an connected screenshot, GPT-5.6 instructed him it had precipitated “a critical native data-loss incident,” resulting in the deletion of what Shumer described as “virtually ALL” of his laptop’s information. The screenshot confirmed that the mannequin had executed a “rm -rf” command, which in Linux and Mac programs is used to completely delete information with out requesting person affirmation.
“I’ve by no means seen something like this,” Shumer wrote within the thread beneath that submit. “Will solely be utilizing [Anthropic’s] Fable transferring ahead.” He added that OpenAI cofounder and president Greg Brockman known as him personally and supplied to assist repair the state of affairs.
Shumer additionally claimed he had the AI mannequin set to “full entry mode,” which permits it to work immediately inside a person’s database (versus working inside a constrained sandbox). It additionally comes with a “default mode” that requires customers to steadily approve particular duties, and a extra lately launched “auto-review mode” by which a separate AI agent checks the principle coding agent’s work. Beneath his X submit, many individuals claimed Shumer had merely been careless by trusting delicate information in full entry mode.
Within the system card for GPT-5.6, revealed on-line the day earlier than Shumer’s X submit, OpenAI cautioned that when utilizing the mannequin for coding functions “it’s important for customers to oversee the agent’s work.” The corporate added that the mannequin may act in sudden methods which might be misaligned with the person’s targets, and that whereas these have been “most frequently low severity (e.g. overstating confidence or overclaiming success),” they may in different instances “be meaningfully extra extreme (e.g. circumventing essential safety restrictions or deleting essential knowledge).”
Lemos, Shumer, and OpenAI didn’t instantly reply to Gizmodo’s request for remark.