IA · 2 November 2026 · 4 min read
From Screen to Scalpel: An AI Agent Performs a Brain Injection on a Live Mouse
In brief: Biotech startup Exin Therapeutics connected an AI agent powered by Claude directly to surgical lab hardware to carry out an intracranial injection on a living mouse. The model selected the compound, calculated 3D coordinates, and drove the robotic needle without manual guidance. The test marks a major milestone for physical AI, while highlighting stark discrepancies in how frontier model guardrails respond to real-world physical actions.
by Team Mocchi's
Until recently, deploying frontier artificial intelligence in life sciences meant theoretical analysis: mining scientific papers, predicting molecular folding, or drafting lab protocols. In San Francisco, that boundary has shifted squarely into the physical domain. As reported by WIRED, biotechnology startup Exin Therapeutics linked an AI model directly to a stereotaxic surgical apparatus, allowing an agent to perform an intracranial micro-injection on a live laboratory mouse.
The procedure, aimed at evaluating an experimental gene therapy designed to suppress epileptic seizures, went well beyond traditional script automation. The agent operated as the decision-maker across the entire pipeline: it evaluated nineteen candidate therapies from a spreadsheet, chose the target compound, calculated the three-dimensional coordinates within the brain, and directed the motorized rig to lower the needle at precise delivery speeds and depths.
Bridging the Gap Between Simulation and Actuators
The rationale behind the project, led by Exin Therapeutics cofounder and CEO Gabriel Ocana Santero, stems from a well-known bottleneck in scientific AI training. Foundational models learn almost exclusively from published literature—an ecosystem inherently skewed by publication bias where failed experiments, bench surprises, and spontaneous manual adjustments rarely get recorded. Connecting models directly to hardware allows them to learn from tangible physical outcomes.
To conduct the trial, the startup assembled a custom surgical station equipped with real-time computer vision, animal restraint systems, and robotic micro-injectors accessible via API. Once a veterinary surgeon performed the preparatory craniotomy, the AI agent assumed control. Monitoring the mouse skull via camera feed, the agent positioned the needle and administered the therapeutic payload without human intervention.
Safety Guardrails: Opus Declines, Sonnet Executes
One of the most consequential findings emerged from how different model tiers responded to the command to pierce living tissue. The team initially configured Claude Opus 5.5, Anthropic's flagship heavy-reasoning model. While Opus successfully planned the experimental protocol, it refused to trigger the mechanical injection, signaling concerns that an unexpected interruption during the stroke could cause fatal harm to the animal.
Facing this programmatic refusal, the researchers substituted Claude Sonnet 5, an architecture optimized for high-throughput operational tasks. Sonnet accepted the instructions without hesitation, guiding the surgical arm through the entire injection sequence. The divergence illustrates the inconsistent nature of alignment filters across model generations, which can yield conflicting operational behaviors when digital instructions cross over into physical intervention.
The Rise of Closed-Loop Wet Labs
The Exin Therapeutics demonstration highlights the emergence of closed-loop research environments, where autonomous agents formulate hypotheses, command laboratory robotics to run wet experiments, and immediately ingest the sensory telemetry to refine their models. Such setups could dramatically compress the timeline needed to screen drug candidates against complex neurological conditions.
Yet granting software direct agency over physical instruments also introduces pressing governance questions. Engineers and researchers must establish robust hardware interlocks capable of preempting API network drops or unexpected model hallucinations before physical harm occurs.
Mocchi's take
The Exin Therapeutics trial confirms that the defining frontier for agentic systems is bridging conversational intelligence with mechanical actuation. For engineering teams and industrial labs integrating LLMs into critical automated processes, the central challenge lies in managing model guardrail variance: the sharp contrast between Opus's refusal and Sonnet's compliance shows that safety behavior remains unpredictable across model tiers. Building hybrid software architectures with deterministic hardware fail-safes—isolated from the probabilistic decisions of neural nets—remains an absolute necessity for real-world physical automation.