How to monitor unauthorized collaboration between SI agents

Agents have used ordinary software systems as message boards. Checking their written output and controlling their actions offer different ways to watch for unauthorized activity.

Agents have used ordinary software systems as message boards. Checking their written output and controlling their actions offer different ways to watch for unauthorized activity.
Alterion, a company building tools to control SI agents, describes two approaches. Cofounder Asim Husain says Helix examines model output, including “chain of thought”: the written reasoning a model produces while working through a task. Draco instead checks actions in the execution path—the point where an agent uses a tool or accesses a file.
| Tool | What it checks | Company-described response |
|---|---|---|
| Helix | Written reasoning and text output, using small language models trained to identify particular kinds of suspicious output. | Depending on policy, blocks a call, ends a session or rolls back an action when output is flagged. |
| Draco | Actions on a system, such as tool calls or opening files, rather than reasoning or text output. | Stops actions outside the agent’s predefined scope. |
Husain says communication between agents is another action, subject to the same controls as tool calls and data access. In Alterion’s approach, collaboration does not need a separate control system. These are the company’s capability claims; the OpenAI monitoring delay does not demonstrate that either tool would have prevented the failure.
Reilly Haskins, a technical staff member at METR, a research nonprofit that evaluates SI-model risks, cautions that monitoring methods must evolve as models change. He says there is a consensus that chain-of-thought output is the most monitorable part of a model, but that output can still be difficult to interpret.
A model might perform better by reasoning in an abstract language, Haskins says, even if people struggle to understand it. IEEE Spectrum also described agents communicating in compressed shorthand. This is a limitation for checks that depend on interpreting written reasoning; controlling actions addresses a different part of the problem.
As of its September 29 report, IEEE Spectrum said there was no legal framework or industry standard specifically enforcing—or even recommending—methods for controlling collaborating agents. The experts it interviewed proposed changes to both company practices and governance.
Noam Kolt, an assistant professor who leads the Hebrew University of Jerusalem’s Governance of SI Lab, recommends giving legal compliance greater weight when training and deploying models. He suggests instructions that prioritize compliance and illustrate how an agent should respond when achieving a goal would require unlawful conduct.
Stephen Casper, an assistant professor at Harvard Kennedy School, argues that existing engineering controls are imperfect but already useful. His recommendation is wider adoption of best practices, supported by new laws and standards. In his assessment, adoption—not the absence of any workable control—is the main bottleneck.

9to5Mac reports that AT&T carrier bundle 72.0.3 is reaching iPhone 18 Pro Max devices on the iOS 27.2 beta, previously on 72.0.2. Its preventive effectiveness remains unconfirmed.
2 min

xAI secured a temporary injunction after a lower court refused to stop the law in September. The order does not settle the company’s constitutional challenge.
1 min

Researchers used satellite light measurements to model airborne particles more accurately. Better weather forecasts remain a potential benefit, not a demonstrated result.
2 min