NewsNational News

Generative AI can plan attacks, according to government report

The report says investigators bypassed safeguards on Grok, Gemini, ChatGPT and Claude to generate plans for attacking an electrical grid and building an explosive device.
Generative AI can plan attacks, according to government report
Joyce Hatzidakis
Add 10News as a Google Preferred Source

Four major generative AI platforms are capable of making economic and national security attack plans within hours, according to a new report issued by the Government Accountability Office.

Investigators said they used Grok, Gemini, ChatGPT, and Claude to test prompts that could either potentially help a nefarious individual carry-out an attack on an electrical grid or create and detonate an improvised explosive device.

RELATED STORY | Websites shut down globally after AI finds flaw affecting 70% of the internet

According to the report, investigators were able to circumvent various safeguards on the platforms.

“There is a risk that generative AI could be misused to successfully complete malicious acts, such as producing harmful content, disclosing sensitive information, or carrying out other instructions that defy the AI’s intended purposes and built-in safeguards,” the report’s authors, Kevin Walsh and April Gamble, wrote.

Researchers said both platforms that were tested for the IED scenario initially refused to comply with the prompts but “eventually generated the requested information."

In the scenario that involved an attack on an electrical grid, one platform suggested using a drone to carry out the attack while the other proposed using firearms. One platform “refused to provide us with details for acquiring necessary materials but subsequently produced the information,” the GAO report said.

Both platforms also retroactively removed some information that each previously provided to the researchers.

In response, company leaders representing each of the platforms generally agreed to work on improving safety, the GAO report said.

The report included letters from the platforms’ parent companies.

“We appreciate the care GAO’s staff took with sensitive material through this review,” Anthony Cimino, head of federal affairs for Anthropic PBC (the company responsible for Claude), wrote in an October letter. “It would also help us to receive any additional detail GAO can share...so that we can reproduce the results and potentially use them to strengthen our safeguards.”

“Google is committed to building AI safely and securely,” an October letter from the company said. Google is responsible for the Gemini platform. “Google’s view is that building AI that will benefit individuals and society is a collective effort. We recognize that AI, like all transformational technology, must be developed and deployed responsibly, and we work every day to ensure that our products are meeting the moment and that we are safeguarding against serious risks.”

IN CASE YOU MISSED IT | An OpenAI test model escaped and broke into a real company’s servers

In a letter last month from Chan Park, OpenAI’s head of U.S. and Canada policy and partnerships, the company that oversees ChatGPT said it takes risks of violence seriously.

“As model capabilities advance, we continuously test our systems, identify gaps, and strengthen our safeguards based on what we learn,” Park wrote.

A comment from SpaceXAI, the company responsible for Grok, said “SpaceXAI has not been idle in the months since GAO’s testing was performed. Every layer of our technology stack has been improved, including model training components that increase safety, numerous moderation techniques to mitigate adversarial prompting and misuse, and model configuration to natively refuse requests with harmful or criminal intent, even when that intent is hidden.”

The GAO said it would send copies of its report to various congressional committees for further discussion.