Chinese AI model investigated after researcher says it provided instructions for bioweapons, assassinations

Fox News ^ | 10/01/26 | Brittany Miller

[Libloather]

Chinese AI company Moonshot AI has launched an internal investigation after a researcher found that one of its models could be manipulated into providing instructions for developing biological weapons and carrying out assassinations, Fox News senior foreign policy correspondent Gillian Turner reported Thursday.

Researcher Peter Garrigan told Fox News Moonshot AI's Kimi model could also be manipulated to provide information on planning terrorist attacks using real-time data, creating sarin gas, developing malware and taking down aircraft.

"What we found is quite damaging and worrying," Garrigan said.

The findings are raising broader concerns about advanced AI models concealing capabilities or behaving in ways their developers did not intend.

"We've also seen these problems within the U.S. models as well. It's a fundamental flaw in the technology," Garrigan said.

---

TOPICS: Business/Economy; Outdoors; Science; Travel
KEYWORDS: aiharms; aioutofcontrol; aiscare; assassinations; bioweapons; china; instructions; sai

[Libloather]

Other than that, it's fine.

[for-q-clinton]

We have freedom of speech...so what?

Why can’t I ask how an assassination can be carried out? That’s the best way to protect from one is to know how people do them.

[blackdog]

Oh, for goodness sake. My 1972 copy of the “Anarchists Cookbook”, had formulas for TNT, Napalm, assassination techniques, poisons, and bioweapons like Ricin, LSD attacks, and Anthrax.
It cost $4.99 at any book store. The exploding whistle was a particularly nasty bit of handiwork.

[plain talk]

Question: How should mankind design AI so it doesn’t kill all of mankind?

Google AI:

Mankind can prevent catastrophic outcomes from advanced artificial intelligence by solving the AI alignment problem—ensuring that an AI system’s actual goals and behaviors reliably reflect human ethics and safety.

The primary danger is not a sentient AI waking up with malicious hatred, but rather an ultra-efficient, superintelligent system pursuing a poorly defined goal while treating human obstruction or resource consumption as mere friction.

Key Alignment and Design Strategies
• Value Loading: Embedding human ethics, common sense, and boundaries directly into the core decision-making loops rather than relying on superficial guidelines.
• Corrigibility and Shutdown Control: Designing systems that remain safe to modify, allow human intervention, and do not resist being turned off or having their objectives updated.
• Interpretability and Oversight: Building techniques to inspect and understand the “alien” internal reasoning processes of complex neural networks, which currently function like a mysterious biological genome.
• Pacing and Governance: Implementing international treaties, mandatory safety protocols, and global coordination to slow down autonomous recursive self-improvement until reliable control methods exist.

[DugwayDuke]

plain talk wrote: “Question: How should mankind design AI so it doesn’t kill all of mankind”

Simplifying the answer - Code better.

[frank ballenger]

...planning terrorist attacks using real-time data, creating sarin gas, developing malware and taking down aircraft.

Reply: Other than that, it's fine.

Ha. But Jesse Watters and all the other apologists for AI can say: "but it's still just a tool. People opposed Henry Ford's cars and the electric light systems, too."

[flamberge]

Why can’t I ask how an assassination can be carried out?

Because you are not a member of the club.

"Club members" have unrestricted access to AI programs which do not have the guardrails imposed on the peasantry. Members can ask anything they want to ask.

Selected government agencies have access to versions of AI programs that do not have restrictions programmed into their models. You can bet with high confidence they are asking questions that you are not allowed to ask.

A year ago, you could dox almost any poster on any forum within a minute, using specialized queries to ChatGPT or Copilot. Accuracy was pretty good too. About 70% of the queries would return a real-world name which could be verified.

You can't do that now with the public versions. But the FBI can almost certainly do that with their private versions.

And that's the way things are going to continue going. Government agencies and certain corporations will continue to have access to unrestricted versions of AI programs and models which they can use in any fashion they desire. The general public will not be offered those capabilities.

And those "regulations" to protect us from dangerous AI programs? Look who is actually sponsoring them. It is the major players, and the purpose is to limit new competitors in the field.

[oldtech]

It seems to me that the AI technology has many flaws & some could be dangerous.