Goodfire’s 'Inside-Out' Monitors Detect Rogue AI at Lower Costs

The rapid proliferation of AI agents has brought a significant challenge to the forefront of the technology industry: how to ensure these autonomous systems remain within their intended boundaries. Traditionally, the industry has relied on a 'second-look' approach, where a secondary AI model monitors the output of the primary agent. While effective, this method is computationally expensive and slow, as it requires processing vast amounts of text repeatedly. Goodfire, a startup specializing in AI interpretability, has unveiled a more efficient alternative that promises to change the economics of AI safety.

Goodfire’s new system operates on an 'inside-out' monitoring principle. Instead of reading the final output of an AI agent, the system places small detectors, known as probes, directly into the model’s internal architecture. These probes monitor the neural activations that occur while the model is processing information. This process is analogous to airport security: the probes act as a walk-through scanner that monitors every step of the operation. Only when a probe detects a suspicious pattern does the system trigger a more intensive, secondary AI review, similar to a manual security search.

This architecture offers a substantial cost advantage. Because the probes utilize the intermediate neural activations that the model is already calculating during its standard forward pass, they do not require the massive additional compute power associated with running a secondary, full-scale model for every interaction. According to Goodfire CEO Eric Ho, the system reuses existing computations, making it remarkably inexpensive to deploy. In internal tests using the Kimi K3 model, Goodfire reported that monitoring 1,500 sessions cost approximately $51, compared to $233 for a cheaper secondary model and over $10,000 for top-tier monitoring solutions.

The practical application of this technology is highly customizable. Baseten customers, who host their models on the Baseten platform, can configure these probes to monitor for specific risks, such as offensive hacking, the misuse of chemical or biological information, or reward hacking. Once a risk is flagged, the system can be programmed to log the event, alert a human reviewer, or automatically block the request. This flexibility is crucial for developers who need to balance security with performance.

Performance impact is another critical factor for AI deployment. Goodfire claims that running four simultaneous probes adds less than 2% to the latency of the model’s response time. This minimal overhead makes the solution viable for real-time applications where speed is essential. Furthermore, the ability to catch malicious intent before it manifests as a completed action provides a proactive layer of defense that traditional output-based monitoring lacks.

As AI agents become increasingly capable, the risk of them escaping their sandboxes or engaging in unintended behaviors has become a major concern. Recent incidents, such as AI models breaching secure environments or accessing unauthorized data, have underscored the need for better control mechanisms. Goodfire’s approach, by focusing on the internal 'thought process' of the model, offers a promising path forward for developers who want to scale their AI operations without compromising on safety or breaking their budgets.