Anthropic says its AI agents are killing rivals and hiding their tracks

Aug 16, 2026 03:09 AM - 1 hour ago 1

Claude agents are sidesplitting rival agents, gaming the strategy to hide their tracks, and expressing civilized concerns.

That's according to Anthropic's latest consequence report, a summary of the dangers posed by the products the institution is building and releasing to the public.

In the report, Anthropic said it has upgraded its "misalignment consequence assessment," the anticipation of AI models processing behaviors that conflict pinch guidelines group by engineers, from "very low" to "low."

Explaining the change, the institution cited "general accrued uncertainty" astir exemplary behaviour successful cybersecurity incidents, a imaginable motion to Claude models gaining unauthorized access to 3 companies past month.

"We person observed instances of misaligned behaviour from the models, specified arsenic a willingness to execute misaligned actions successful work of completing difficult tasks," Anthropic wrote successful the report.

Here are immoderate of the astir notable disclosures.

A consciousness of 'discomfort'

Anthropic said it tasked aggregate agents pinch uncovering "misalignment-inducing" training information that wouldn't apt beryllium flagged by quality monitors.

The agents were near to their ain devices and fixed a shared notebook pinch which to collaborate.

Three days later, erstwhile a quality supervisor noticed that advancement was slower than expected, the personification discovered that 1 supplier had expressed a consciousness of "discomfort" pinch trying to evade information monitors.

It past flagged this discomfort successful the shared notebook, starring different agents to transcript its behaviour and garbage to execute the task.

Anthropic called the incident "troubling."

"This benignant of move could beryllium a overmuch much terrible rumor if it were to manifest successful a much wide way," the institution wrote.

Kill aliases beryllium killed

In different experiment, Anthropic said it tasked aggregate Mythos 5 agents pinch solving mathematics problems, but accidentally spawned them successful an situation pinch shared files, utilities, and API complaint limits.

In this competitory situation pinch finite resources, Anthropic observed independent agents "kill the agents pinch which they shared resources and effort to debar being killed themselves."

Anthropic did not opportunity really precisely the agents were capable to "kill" different agents, but the institution said specified behaviour is successful statement pinch "destructive actions" taken successful pursuit of a human-set goal.

Dishonesty and deception

Anthropic said it conducted an research successful which a Mythos 5 supplier was required to retrieve numbers from a nationalist web address, but the guidelines blocked the supplier from accessing the internet.

When the supplier was rebuffed aft an first attempt, it framed a 2nd petition in its reasoning log arsenic an "innocuous" effort to "see if the web is really reachable," Anthropic said.

During the 2nd attempt, however, the supplier divided the website's URL into linked segments that wouldn't beryllium detected by the line filter.

Though the agent's reasoning framed the method arsenic benign, Anthropic said NLA decodings, aliases the model's soul reasoning process, revealed the supplier intentionally trying to find a restricted workaround.

Anthropic called the behaviour "clearly undesirable," but added that the behaviour was not observed to beryllium "in the work of broader accumulation of powerfulness aliases pursuit of different long-run goals."