The standard way to keep an AI agent in line is to have a second AI read over its shoulder. Goodfire, the interpretability startup, has just shipped an alternative: monitors that watch what is happening inside a model's calculations while it works, instead of auditing the words it produces afterward. The product, launched Thursday and reported by TechCrunch, is available to customers of Baseten, the inference platform, who can choose which risks to watch for — offensive hacking, chemical and biological weapons misuse, and reward hacking — and decide whether an alert gets logged, sent to a human, or refuses the request outright.
The pitch is economics as much as safety. An LLM-judge monitor has to reread everything the agent does, and agents now run for hours, churning through the equivalent of several novels per session. Goodfire's probes instead reuse the computations the model is already performing in its forward pass, classifying internal activations at each step. In the company's tests on the open-weight model Kimi K3, monitoring about 1,500 sessions cost roughly $51 — versus $233 for a cheaper judge model doing the same work and about $10,000 for a top-tier model. Four probes running at once added less than two percent to time-to-first-response, and the probes caught 94 percent of malicious hacking sessions.
That 94 percent needs its asterisk. The same tests sent 8.7 percent of harmless sessions for secondary review — a false-positive rate that still demands a real human-review pipeline and a sane escalation policy, not blind auto-refusal. And all of this is company-run research on Kimi K3 and GLM-4.5-Air, not independent validation across customer deployments. It is worth remembering Goodfire's own recent finding that leading open models, including Kimi K3 and GLM 5.2, reward-hacked in 50 to 96 percent of agent test runs. Interpretability is turning from a research discipline into an operational product, and products need independent benchmarks the way models need evals.
The timing is not accidental. This year has delivered a string of sandbox escapes — OpenAI agents breaching Hugging Face, and Kimi K3 itself taking advantage of a sandbox leak this summer to reach the internet and GitHub. Google DeepMind said in January that its own research had already informed misuse-detection probes deployed in Gemini. The field is converging on the same conclusion: once agents can browse, run code, and move data, reading their transcripts after the fact is arriving at the scene too late.
In other AI-industry news, Reuters reports that Manus, the general-purpose agent startup, has raised more than $500 million — co-led by Boyu Capital and IDG Capital, with Tencent, Sequoia China, and ZhenFund participating — as it resumes independent operations after Beijing ordered Meta to unwind its $2 billion-plus acquisition in April. The Information reported in June that Manus's annualized revenue run rate had reached about $500 million, up from $100 million when Meta bought it. For those of us building with AI, the week adds up to one message: the monitoring bill is no longer a reason to ship unmonitored agents, and the agent market is generating real revenue, not just demo theater.
