Breaking

Goodfire Launches Inside-Out Monitors for AI Agents on Baseten

October 8, 2026 at 12:10 EDT

AI Agents
Security
Foundation Models

Goodfire launched activation-based monitors for AI agents on Thursday, pitching them as a lower-cost alternative to having another LLM review every step an agent takes. The system reads internal model activations while an agent runs and escalates suspicious cases to a separate judge only when a probe fires.

Oct. 8, 2026 · Goodfire

Inside-out monitors for AI agents

Goodfire's new activation probes watch the model while it works, then escalate only suspicious runs to a heavier AI judge. The pitch is lower cost, lower latency, and earlier warning for open-model agents.

Cost to monitor about 1,500 Kimi K3 sessions

$51

Activation probes

$233

Cheaper AI judge

The stacked columns show the lower-cost comparison: one block is about $46, so $51 is about one block and $233 is about five blocks.

94%

Malicious hacking sessions caught

8.7%

Harmless sessions escalated

<2%

Added response-start latency with four probes

$10,000

Top-tier judge baseline

How the cascade works

1. Agent runs

The model starts a tool-using task.

2. Probes read

Internal activations are checked as they are produced.

3. Flag if needed

Only suspicious runs move to heavier review.

4. Escalate

A judge, human, or policy can stop or approve the action.

Why teams care

Always-on guardrails become cheaper, and internal signals may appear before a risky action reaches text or tools.

What still needs work

Probes need model-internal access, calibration, labeled data, and revalidation after some model updates.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.