Goodfire’s new monitors inspect a model’s internal activity while an AI agent works, rather than relying only on the agent’s outputs. Small probes run during generation, and when a probe flags a concern, a separate AI model can review that case. TechCrunch reported the launch on October 8, 2026, and says the monitors are available to Baseten customers. The cost and detection figures come from Goodfire’s own tests, and no independent replication of these launch-specific results has been established. They are best read as the company’s numbers for a specific setup, not as a general measure of performance.
How the probe-first design works
The monitor is built as a cascade. A cheap detector runs on every step, and an expensive reviewer is called only when the detector raises a flag. In the sequence reported for Goodfire’s approach:
As an Amazon Associate I earn from qualifying purchases.
- The model processes a step of the agent’s task and computes internal activations, as it does during every forward pass.
- A lightweight probe reads those activations and scores whether they match a risk signal it was set to watch for.
- If the score crosses the probe’s threshold, the case is escalated to a separate AI model that reviews it.
- The flagged case then receives whichever response the customer has configured.
The distinction from output-based monitoring is what the monitor can see. A conventional second-model monitor reads what the agent produced or did. A probe reads signals inside the model that produced it. TechCrunch’s report does not describe how the probes are trained or which layers they read, so the mechanism is described here only at the level the report supports.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Why the probes are cheap to run
Goodfire CEO Eric Ho puts the cost argument this way: “Internal activation monitors are really cheap because they reuse the computations in the forward pass.” The probe does not need a second model to read the transcript. It reads signals the model already computes. The expensive reviewer then runs only on flagged cases, which is why escalation rate, not step count, becomes the main driver of review cost.
#1 Best Overall
- CRISP CLARITY: This 23.8″ Philips V line monitor delivers crisp Full HD 1920x1080 visuals. Enjoy movies, shows and videos with remarkable detail
- INCREDIBLE CONTRAST: The VA panel produces brighter whites and deeper blacks. You get true-to-life images and more gradients with 16.7 million colors
- THE PERFECT VIEW: The 178/178 degree extra wide viewing angle prevents the shifting of colors when viewed from an offset angle, so you always get consistent colors
- WORK SEAMLESSLY: This sleek monitor is virtually bezel-free on three sides, so the screen looks even bigger for the viewer. This minimalistic design also allows for seamless multi-monitor setups that enhance your workflow and boost productivity
- A BETTER READING EXPERIENCE: For busy office workers, EasyRead mode provides a more paper-like experience for when viewing lengthy documents
Sensitivity is a threshold setting
In Goodfire’s September 17, 2026 reward-hacking research, the company describes activation probes as adjustable: sensitivity changes when the threshold changes. A lower threshold catches more suspicious activity and sends more cases to review. A higher threshold reduces escalations and may miss more. The tested trade-off between these settings is part of the company’s reported results, and the report does not state a recommended threshold for any deployment.
What CTO Dan Balsam means by catching things earlier
Goodfire CTO and co-founder Dan Balsam says, “The great advantage is that you can catch things before they happen.” Because the probes run during generation rather than after an output is complete, the intended benefit is earlier flagging. The reported tests do not measure how often earlier flagging changed an agent’s final action, so treat this as the design goal rather than a demonstrated outcome.
What Baseten customers can configure
TechCrunch reports that Baseten customers can select from several risk categories and several responses. The report does not describe the configuration interface or whether the categories can be combined in custom ways.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
- CRISP CLARITY: This 22 inch class (21.5″ viewable) Philips V line monitor delivers crisp Full HD 1920x1080 visuals. Enjoy movies, shows and videos with remarkable detail
- 100HZ FAST REFRESH RATE: 100Hz brings your favorite movies and video games to life. Stream, binge, and play effortlessly
- SMOOTH ACTION WITH ADAPTIVE-SYNC: Adaptive-Sync technology ensures fluid action sequences and rapid response time. Every frame will be rendered smoothly with crystal clarity and without stutter
- INCREDIBLE CONTRAST: The VA panel produces brighter whites and deeper blacks. You get true-to-life images and more gradients with 16.7 million colors
- THE PERFECT VIEW: The 178/178 degree extra wide viewing angle prevents the shifting of colors when viewed from an offset angle, so you always get consistent colors
- Risks: offensive hacking, chemical and biological weapons misuse, and reward hacking.
- Responses: logging, human review, or refusal.
Three approaches compared on the reported Kimi K3 tests
TechCrunch’s October 8, 2026 report gives the figures below for Goodfire’s Kimi K3 tests, which covered about 1,500 sessions. Cells marked “not stated” are absent from that report, and the report does not name the two comparison models.
| Factor | Goodfire probes with reviewer on flagged cases | Cheaper model monitoring every step | Top-tier model monitoring |
|---|---|---|---|
| What is read | Internal activations at each step; a separate AI model reviews flagged cases | Every step; whether it reads outputs or internals is not stated | Not stated |
| Reported cost for about 1,500 sessions | About $51 | About $233 | About $10,000 |
| Reported detection and escalation | 94% of malicious hacking sessions caught; 8.7% of harmless sessions escalated | Not stated | Not stated |
| Reported added latency | Less than 2% more time to first response with four probes (company figure) | Not stated | Not stated |
| Independent validation | Not established | Not established | Not established |
What the cost gap does and does not show
Dividing the reported totals by 1,500 sessions gives roughly $0.03 per session for the probe approach, $0.16 for the cheaper model on every step, and $6.67 for the top-tier model. The probe approach costs about 22% of the cheaper-model figure and about 0.5% of the top-tier figure. These are simple divisions of TechCrunch’s reported totals, not separately measured costs.
The gap is large enough that the direction of the claim is plausible from the mechanism alone. Its size, however, depends on conditions the report leaves out: the model names, whether all three approaches ran on identical tasks and settings, and how much of the $51 went to reviewer calls on the flagged sessions.
Rank #3
- Clear visuals. Fluid motion: A 144Hz refresh rate and 1ms MPRT deliver smooth, tear‑free motion across work, gaming, and streaming for clearer, more fluid viewing.
- Eye comfort: TÜV Rheinland 3‑star* certification reduces harmful blue light while preserving stunning color quality without compromise. *TÜV Rheinland 3-star eye comfort certification.
- Wide viewing angle: Get consistent views across a wide 178° /178° viewing angle.
- In-Plane Switching (IPS): See excellent color accuracy and consistency across wide viewing angles with In-plane Switching (IPS) technology.
- Ultra-thin bezels: Maximize your viewing experience with thin bezels.
Detection and false escalation
The 94% figure refers to malicious hacking sessions, but the report does not say how many malicious sessions were in the test, so the figure rests on an undisclosed sample size. The 6% that were missed are not characterised in the report. The 8.7% escalation rate applies to harmless sessions. Put in practical terms, roughly one in every 11.5 harmless sessions reached the reviewer. In a deployment dominated by benign work, reviewer cost would scale with that rate, so the escalation rate on your own traffic matters more than the headline cost figure.
Recommended Free Tools
Latency
Goodfire says that running four probes adds less than 2% to time to first response. The report does not state the hardware, model size, or measurement method behind that figure, and it does not report latency for the reviewer step, which runs only on flagged cases.
What Goodfire’s reward-hacking research measured
Goodfire’s reward-hacking research, dated September 17, 2026, is a separate company study from the launch tests. It reports two results, both attributed to the company’s own experiments:
Rank #4
- CURVED FOR ENHANCED ENGAGEMENT: An immersive viewing experience with a curved monitor that wraps more closely around your field of vision; It creates a wider view, enhancing depth perception and minimizing peripheral distraction
- SMOOTH PERFORMANCE FOR SEAMLESS CONTENT: Stay in the action when playing games, watching videos, or working on creative projects; The 100Hz refresh rate reduces lag and motion blur so you don't miss a thing in fast-paced moments¹
- MORE GAMING POWER: Gain the edge with optimizable game settings; Color and image contrast can be adjusted to see scenes more vividly and spot enemies hiding in the dark; Game Mode adjusts any game to fill the screen so you can view every detail²
- KEEP IT EASY ON THE EYES: Care for your eyes and stay comfortable, even during long sessions; Advanced eye comfort technology certified by TÜV reduces eye strain by minimizing blue light and reducing irritating screen flicker²
- INCREASED VERSATILITY: Connect to more; Plug devices straight into your monitor for increased flexibility, making your computing environment even more convenient
- Reward-hacking rollouts at 50–96% across three open models and three agentic benchmarks.
- A probe-plus-LLM setup on Kimi K3 that cut LLM monitoring cost by 90%, with about a 1% precision drop. Precision here means the share of flagged cases that were genuine problems.
What the internal signal looks like
According to the same research, the internal signal associated with reward hacking includes concepts such as cheating, gaming a metric, and avoiding detection. The company says the probes catch instances that chain-of-thought monitors missed and generalize beyond their training data. It also acknowledges that the work uses simple techniques and that further work may improve performance. These are company findings. Activation signals are not perfect access to intent, and a probe that picks up a concept has not thereby established what the agent meant to do.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Biosecurity monitors: a separate line of work
Goodfire’s biosecurity research, dated October 1, 2026, argues that task text alone can be insufficient for dual-use biology workflows. Its monitors combine protein-model embeddings with task context to assess biological sequences. The company reports evaluation on a custom benchmark and says the approach holds up to paraphrasing and fragmentation. Those are company-reported claims. The report does not say the biosecurity monitors are part of the Baseten offering.
The company’s framing of the broader goal is direct: “We believe it will be important to take a defense in depth approach that screens at both the point of synthesis and the point of design, in particular during agentic workflows.”
Best Value
- 【INTEGRATED SPEAKERS】Whether you're at work or in the midst of an intense gaming session, our built-in speakers provide rich and seamless audio, all while keeping your desk clutter-free.
- 【EASY ON THE EYES】 Protect your eyes and enhance your comfort with Blue-Light Shift technology. This feature reduces harmful blue light emissions from your screen, helping to alleviate eye strain during long hours of use and promoting healthier viewing habits.
- 【WIDEN YOUR PERSPECTIVE】Our sleek minimal bezel design ensures undivided attention. The nearly bezel-free display seamlessly connects in a dual monitor arrangement, delivering an unobstructed view that lets you focus on more at once, completely distraction-free.
What to verify before relying on the claim
The cost argument holds up as a mechanism: a cheap probe on every step with an expensive reviewer on flagged cases should cost less than a reviewer on every step. Whether the reported size of that saving applies to your deployment is a separate question, and the published material does not answer it. Before relying on the figures, a buyer or evaluator should:
- Ask for the number of malicious and harmless sessions behind the 94% and 8.7% figures, and the model names used for the comparison approaches.
- Ask whether reviewer calls on flagged sessions are included in the per-session cost.
- Run the monitor on a sample of your own agent traffic and record the escalation rate, since that rate sets the reviewer load.
- Test the threshold setting at the sensitivity you need, and record what it misses as well as what it flags.
- Confirm current availability, pricing, and supported models directly with Baseten or Goodfire, since the reported material does not establish them.
Availability and open questions
The launch is reported for Baseten customers. Goodfire says it works with inference providers to deploy monitors, but the report does not describe other providers. The sources do not establish public pricing, exact customer eligibility, the list of supported models, contractual service levels, or whether self-serve onboarding exists.
The detection rate is measured on Goodfire’s tests and should not be taken as a rate for other models, risk categories, deployments, or real-world incidents.
Sources: TechCrunch launch report (October 8, 2026); Goodfire reward-hacking research (September 17, 2026); Goodfire biosecurity research (October 1, 2026).
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
