OpenAI Signs Goodfire, Reported $5 Billion-Plus Buyout Target, to Inspect Training Runs
1 day ago / Read about 19 minute
Source:TechTimes

KIRILL KUDRYAVTSEV/AFP via Getty Images

OpenAI said Monday it signed Goodfire, an interpretability startup it reportedly explored buying this summer for more than $5 billion, as one of four outside groups in its first round of embedded safety assessments. Goodfire is paid for the work, and OpenAI's own researchers are its primary audience. Readers weighing whether outside checks on frontier AI can be trusted now have three facts to hold together: a fee-paid assessor, a reported former acquisition target, and a lab that keeps the right to redact what the assessor publishes.

Goodfire works in interpretability. Mechanistic interpretability studies a neural network's internal structures and circuits, which Wikipedia says is similar to reverse engineering software, instead of judging a model only by its outputs. That matters during training because a monitor that reads internals can, in principle, catch behavior that output testing misses.

The Information reported Monday, according to secondary summaries, that Altman and other OpenAI leaders spoke with Goodfire several times this summer about an acquisition or a partnership, and that a senior executive raised a price above $5 billion. That is more than four times the $1.25 billion valuation Goodfire set in a $150 million Series B in February. No deal followed, and the reports say Goodfire's leaders preferred to stay independent. TechTimes found no public comment from either company and could not read the paywalled original, so the buyout talks remain a single-source report.

Goodfire Is Paid, and OpenAI's Own Researchers Are Its Primary Audience

OpenAI says payment to Goodfire and Apollo Research does not depend on evaluation results. METR and Transluce do not accept compensation, though they may accept compute or API credits.

Goodfire's first-phase terms are less detailed than Apollo's. Public summaries reviewed by TechTimes do not state its duration, start date or specific access.

Same Probes That Monitor Training Can Also Steer It, Goodfire's Own Research Shows

Goodfire's published method, Reinforcement Learning from Feature Rewards, uses lightweight probes on internal representations as reward signals for reinforcement learning. In tests on Gemma-3-12B-IT, Goodfire reports a 58% hallucination reduction with its probing harness, at roughly one-ninetieth the cost per intervention of an LLM judge.

The same page says the reward probes run on a frozen copy of the model rather than the model being trained, and the authors call their probe-transfer result only a first data point.

That dual use is the largest question the OpenAI agreement raises. An instrument that reads a model's internals during training is also a candidate training signal, and a monitor used that way stops being a neutral witness. Researchers have shown that obfuscated activations bypass defenses built on probes. Goodfire's tests are small-scale, and TechTimes found no public description of how its OpenAI work will separate monitoring from optimizing.

Skeptics doubt the whole approach. Dan Hendrycks and Laura Hiscott argued in May 2025 that mechanistic interpretability has repeatedly failed to hold up, and noted that Google DeepMind researchers deprioritized sparse autoencoders, which the authors call its leading approach.

Can an Assessor Paid by OpenAI Stay Independent?

As Wikipedia summarizes it, an auditor can never achieve complete economic independence when the audited party sets fees and appointments. Independence has three parts: choosing the approach, unrestricted access to information, and freedom to report findings.

The Goodfire case adds a wrinkle: OpenAI reportedly explored buying the company this summer, then signed it as a paid assessor. Goodfire also criticized OpenAI's process publicly before any of this. In July 2025 it wrote that OpenAI's post-training and black-box evaluation had given it little insight into what it was changing inside a model.

OpenAI Can Redact Findings and Apollo Says Its Pilot Falls Short of Its Criteria

Under the shared terms, assessors decide whether and what to publish, but OpenAI may redact confidential information and may request a delay to fix security or safety problems. Disputes go first to OpenAI's Safety Advisory Group and then to the Safety and Security Committee of the OpenAI Foundation's board.

Apollo says six of its evaluators will work inside OpenAI for a planned four months, and that the pilot does not yet meet all its criteria for full embedded evaluations. One stated criterion is the ability to share findings with an appropriate outside audience. By Apollo's account, OpenAI agreed not to withhold information merely because it is unfavorable and must explain refusals in writing. Apollo did not say when its pilot will begin, according to TechCrunch's report.

The track record matters here. METR and Redwood Research reported on Oct. 22 that their earlier pilot found vulnerabilities in OpenAI's monitoring system, including an exploit that could have hidden hundreds of agent actions. They agreed to publish only a short public summary of a 61-page report, which OpenAI could redact.

Altman Backs Mandatory Evaluations While Goodfire's Ho Has No Public Policy Record

OpenAI has proposed mandatory third-party evaluations of advanced models, diverging from the White House's voluntary framework on at least two points. Altman has also urged lawmakers to fund testing, according to people familiar with his message, rather than require federal sign-off before model releases. TechTimes found no public legislative or regulatory positions from Goodfire CEO Eric Ho.

Anthropic Backed Goodfire Early and Now Pays Accenture for Its Own Evaluators

Anthropic took part in Goodfire's Series A, a $50 million round announced in April 2025, so an OpenAI purchase would have meant acquiring a company a rival helped fund. Anthropic is meanwhile paying Accenture's Faculty unit for embedded evaluation, and each company expects to invest at least $1 billion over five years in this area.

Readers should watch for three details: the full terms of Goodfire's agreement, what OpenAI redacts from Goodfire's first findings, and whether Goodfire's work stays on the monitoring side of the line between watching a model and training it.

Frequently Asked Questions

Did OpenAI try to buy Goodfire?

The Information reported that OpenAI leaders discussed buying Goodfire or partnering with it this summer, with a price above $5 billion raised by a senior executive. No deal resulted, and TechTimes found no public confirmation from either company.

What will Goodfire do inside OpenAI?

Goodfire will apply interpretability, the study of model internals, in OpenAI's embedded assessments, with OpenAI researchers as the audience for its findings. Public summaries do not detail its scope or schedule.

Can Goodfire publish what it finds?

Assessors decide whether to publish, but OpenAI reserves the right to redact confidential information and may ask for delays.

Can a tool that monitors an AI model also be used to train it?

Yes. Goodfire's own research uses probes as reward signals for reinforcement learning, which is why researchers ask whether a monitor used that way remains independent evidence.