Goodfire

Goodfire is an AI interpretability company whose tools read a model's internal activations to monitor, debug and guardrail it. It is for teams that train or serve frontier and open-weight models and need cheaper, faster misuse and reward-hacking detection than running a second LLM as a judge on every step.

Goodfire is a San Francisco public benefit corporation that builds interpretability tools, meaning methods for understanding what happens inside a neural network. Its co-founders include CEO Eric Ho and chief scientist Tom McGrath, who founded DeepMind's mechanistic interpretability team, with Dan Balsam as CTO. The LLM offering covers three jobs. Guardrail is a multi-stage, low-latency system that blocks unsafe agent behaviour such as cyber or CBRN misuse at training and inference time. Detect uses activation and action monitors to catch reward hacking during training. Debug traces regressions back to specific training data, environment bugs or runs. The approach puts small probes on the model's internal signals; only exchanges a probe flags go to an LLM judge. In an October 8, 2026 write-up, Goodfire reported deploying cyber-misuse monitors on Kimi K3 and GLM 5.3 inside an SGLang serving stack at about 93% recall with a 5.5% benign interruption rate, roughly 2 to 42 ms of added latency per turn and under $200 per million exchanges. TechCrunch reported the monitors are offered to customers of Baseten, which formed a safety partnership with Hugging Face and Goodfire in September 2026. Goodfire also has a life-sciences line, and its website lists Arc Institute, Mayo Clinic, Microsoft, Prima Mente, Prime Intellect and Rakuten as organizations using its work. It raised a $150 million Series B at a $1.25 billion valuation led by B Capital on February 5, 2026.

Category

Governance & Security

Pricing

Contact for pricing

Target Market

CTOs, Heads of AI, AI Safety Teams, Inference Providers

Deployment

Self-hosted

Headquarters

San Francisco, United States

• ✓
  Probe-based guardrails

  Lightweight probes read internal activations during inference and escalate only flagged exchanges to an LLM judge, which can refuse the request.

• ✓
  Reward-hacking detection

  Activation and action monitors flag reward hacking during RL training so teams can fix environments, adjust rewards or stop runs early.

• ✓
  Data attribution debugging

  Traces harmful or regressed behaviour to specific training datapoints; Goodfire reported a 63% behaviour reduction after filtering them and retraining.

• ✓
  Serving-engine integration

  Probe scores are computed inside SGLang during serving, so monitoring added under about 2% to throughput in Goodfire's tests.

• ✓
  Configurable risk policies

  Customers choose which risks to monitor and whether a flagged event is logged, sent for human review or refused outright.

• •
  Guardrails for hosted open models

  An inference provider runs cyber-misuse probes on Kimi K3 or GLM 5.3 and blocks harmful agent sessions before tool calls execute.

• •
  Cheaper safety monitoring

  Replace an LLM judge on every turn with probes that escalate only suspicious exchanges, cutting judge cost roughly 29 to 49 times per Goodfire.

• •
  RL training oversight

  Monitor reinforcement learning runs for reward hacking and stop or fix a run early instead of discovering the problem after training finishes.

• •
  Hallucination reduction research

  Use probes on internal representations as reward signals; Goodfire reported a 58% hallucination reduction on Gemma 3 12B without benchmark loss.

Goodfire publishes no list pricing; access starts with a demo request on goodfire.com. Its published monitoring cost is about $65 to $185 per million exchanges in compute and judge cost on Kimi K3, which is a run cost, not a licence price. Baseten customers can reach the monitors through that platform, under terms neither company has disclosed.

This page was written from 8 sources, 5 on domains other than goodfire.com.

1. 1.goodfire.com — languagevendor
2. 2.goodfire.com — our series bvendor
3. 3.goodfire.com — production cyber monitors on kimi k3vendor
4. 4.techcrunch.com — goodfire says its new inside out monitors catch rogue ai age
5. 5.techcrunch.com — base labs launches an open weight ai safety partnership with
6. 6.venturebeat.com — goodfire raises 7m for its brain surgery like ai observabili
7. 7.pulse2.com — goodfire 150 million series b at 1 25 billion valuation rais
8. 8.news.ycombinator.com — item