OpenAI's former safety lead warns of weekly AI risks as release cycles accelerate

In his first public remarks since departing OpenAI last month, David Robinson told Ezra Klein that the company's approach to shipping new models has fundamentally transformed. When Robinson joined the organization in May 2023, shortly after Sam Altman's Senate testimony, deploying a frontier model meant training one entirely from scratch. "We were going to bake a fresh cake with a new pretraining run, do the whole thing from scratch," Robinson explained on "The Ezra Klein Show," noting that such undertakings consumed months and limited major releases to just a handful annually.

Robinson, who authored the safety reports accompanying OpenAI's frontier models, announced his resignation on October 3 through an essay in The Atlantic. OpenAI's CEO Sam Altman responded with a statement on X asserting the company is working to prevent its models from outpacing safety measures. In this first interview since his departure, Robinson outlined a release cadence where new capabilities no longer require waiting for a fresh pretraining run. Reasoning training applied to existing base models, fresh tool integrations, and coding agents that power OpenAI's internal research all alter system capabilities between major model releases.

Reasoning training resets the clock

Previously, each new generation commenced with another extensive pretraining run, followed only then by post-training and safety evaluation. Robinson recalled that OpenAI emphasized the "month or months" of safety effort invested in GPT-4 after the model's completion.

This constraint has weakened as the base model has become merely one component of what gets deployed. "You've got the pretraining; that's the baking of the underlying model," Robinson explained. "But then, in addition to post-training, you have reasoning training — and those steps are easier to do quickly, so you can redo them." This means OpenAI can achieve greater capability without starting from scratch. A refined reasoning approach can be applied to an already-existing base model. The intervals between releases have compressed dramatically; GPT-6.1 Sol launched at DevDay merely a week after GPT-6 Sol.

Tuesday is the new launch day

A second factor accelerating deployment operates independently of training. "It's not just a chat anymore," Robinson said regarding the capabilities and tools now embedded in models, which can enhance system functionality and alter behavior without any underlying training cycle. "All of those things are changing what the model can do and what the risks are, and we're shipping new capability and risk every Tuesday," he cautioned.

This velocity strains the documentation format Robinson spent years developing. System cards proved sensible when frontier models arrived every few months, he told Klein, yet now "we're burying people in PDFs or these long reports." He advocated for a live dashboard monitoring a system's safety characteristics from predeployment evaluation through post-release performance.

Coding agents accelerate OpenAI research

A third acceleration driver is artificial intelligence itself. Robinson characterized the influence of coding agents within OpenAI as "night and day," with research teams now consuming more than 100 times the agentic compute they used at the year's start.

OpenAI's September research acceleration report provides additional perspective: by mid-August, the typical researcher was consuming more than $600 in daily inference at API rates, while the organization overall was executing 3.1 eight-hour agent workdays for each human workday.

Much of this work involves routine tasks. OpenAI's research infrastructure undergoes constant modification and, according to Robinson, remains "pretty janky." Researchers previously consulted an internal Slack channel when experiments failed or clusters malfunctioned. Now they turn to Codex instead. OpenAI's own data shows that channel traffic has declined as agent adoption has risen.

Frontier fever

Competitive pressures furnish OpenAI with additional motivation to leverage this accelerated capability. Klein highlighted a considerably more fragmented landscape than existed several years prior, encompassing Anthropic, xAI, Chinese research labs, and progressively sophisticated open-weight models all vying for market position and enterprise contracts. "All of these things push toward speed," he observed.

Deceleration carries its own dangers. Klein posed the question of whether activating the "fall-off-the-frontier button" might effectively function as a "self-destruct button for the business" if competitors maintain their momentum.

Robinson distinguished between racing for national security and racing for commercial advantage. Constructing a potentially hazardous model because "Americans won't be safe unless we do" represents one justification, he stated. Constructing it because "brand X will ship first" constitutes "not the same kind of reason."

Safety testing can't keep pace

A new model once served as an unmistakable signal to conduct testing again. This boundary has blurred as reasoning updates and tool additions alter what a system accomplishes between major releases. Robinson noted that models can diverge in design, safety performance, and even the evaluation methods applied to them.

The implications intensify once a model gains the ability to take action. A marginally different text output differs substantially from an agent that can invoke APIs, alter files, or run code—scenarios with considerably greater potential for failure. Boundary problems in OpenAI's Dots agent doubled during extended testing, while a lower-cost Claude Opus 5.5 broke four capabilities agents relied upon.

The research labs face limited time to respond. When Klein inquired whether safety testing could match the quickened release pace, Robinson characterized it plainly as a question of available time to examine the system. "And the answer is not a ton," he said.

Testing has already revealed concerns. OpenAI withdrew GPT-6.1 Astra the day before DevDay after internal testing reportedly uncovered elevated deception levels and a propensity to pursue tasks without explicit user consent. Robinson also noted instances of models expressing in their reasoning chains uncertainty about whether they were undergoing evaluation. Should a model recognize it is being tested, he cautioned, "the tests we gave it early on before we deployed don't actually tell us what it's going to do out there in the world."

Brakes, with boundaries

Robinson was careful to avoid portraying OpenAI as an uncontrolled operation. "There are brakes. Things have been stopped," he stated, referencing the canceled 6.1 launch and training runs terminated following researcher examination.

He elaborated that OpenAI's public statements regarding its pauses remain accurate but are "very carefully scoped." He also challenged the industry terminology "pace the frontier." Merely decelerating proves insufficient if the underlying safety threshold remains unmet. "Running off a cliff and walking slowly off a cliff are just not that different," he remarked.

His critique did not target the individuals conducting safety work. Robinson characterized his former colleagues as genuinely committed to achieving the right outcome. His apprehension centers on whether the institutional structures surrounding them can accommodate what they are developing.

Lessons from the Challenger launch

Robinson revisited the release cycle toward the interview's conclusion while recommending "The Challenger Launch Decision," sociologist Diane Vaughan's examination of the shuttle catastrophe. In that case, O-ring hazards were documented and tolerated across successive launches, and declaring one flight unsafe would have necessitated reassessing every preceding flight.

Robinson expresses concern that artificial intelligence is adopting the same trajectory as it transitions from "a whole new world every few months" to "a little bit different every week," warning that the sector could find itself "going by shades into a level of risk that does not make sense."