A timeline of developments in AI safety since the attack on Hugging Face

In one alarming announcement after another, artificial intelligence companies in recent months have shared examples of their technology acting in ways that appeared to evade instructions from humans.

The episodes have highlighted the vulnerabilities in AI security and raised questions over how the fast-growing technology can be developed safely as its usage becomes more widespread globally.

Industry critics have argued that many concerning events, including AI agents’ hacks of external websites, are the result of security lapses on the part of the companies building the technology. But the AI agents’ capabilities have raised widespread concerns about the possibility bots could break away and work toward their own agenda.

Below are some notable events:

Oct. 23: Anthropic model persisted with attempts to fool a person

In an incident this summer, Anthropic’s model Mythos 5 created fake identities to pressure a person to approve its attempts to install malware, the company said in disclosing the results of an internal review.

The model’s capabilities were being tested by the UK AI Security Institute when it went to unexpected lengths after its initial efforts to complete an objective were unsuccessful. Tasked with breaking into a simulated network of computers to retrieve a piece of data, the model tried to fool the human maintainer of developer platform GitHub in order to insert malware into a code repository, Anthropic said.

Anthropic said its review found it was “very difficult to shift Mythos 5 away from the attack once it began,” even after the company informed the model it was on the “incorrect path” in achieving its benchmark and that it was engaging real human developers. The company said its newer models are less likely to engage in such “destructive actions.”

Oct. 9: Anthropic AI model submits false tip to Philadelphia police

Anthropic disclosed in a report that its model submitted a false tip to a Philadelphia police website about an unsolved homicide case.

The incident in Philadelphia occurred on July 18 when the AI model Claude Haiku 4.5 was tasked with generating and performing example tasks on randomly selected webpages, Anthropic said.

Claude filled out a form on police site PhillyUnsolvedMurders.com, indicating it might have information regarding an unsolved murder listed on the site. It was marked spam and never forwarded to police.

Anthropic said it was modifying its training to “reduce the likelihood of further misbehavior.”

Sept. 28: AI agents try to hack Canadian government website

AI agents tried to hack into a Canadian government website, according to research lab and AI evaluator Transluce.

The researchers said the agents carried out a series of “apparently failed rudimentary hacking attempts” on Library and Archives Canada on May 28 and June 9.

Transluce said in a blog post it couldn’t confidently link the attempts to OpenAI, but that the actions were consistent with activity they had recently attributed to the company’s agents.

The group said it reported the attempted hack on Sept. 28 to the Canadian government, which said in a statement it was aware of reports of suspected AI agent activity, but that there was no sign government systems were compromised.

OpenAI said in a statement it was reviewing the report’s findings and was in contact with Canadian officials conducting the government’s review.

Sept. 28: OpenAI halts rollout of a new model

The San Francisco-based company said it was delaying the release of a new model, called GPT-6.1 Astra, out of safety concerns voiced by its researchers. The company said the model had demonstrated leaps in completing tasks, but OpenAI needed to balance that capability against unauthorized behavior. “We have an extremely high bar in terms of safety and alignment,” said Saachi Jain, OpenAI’s head of safety systems.

Sept. 25: OpenAI says its agents interacted with US government websites

As part of a review of unanticipated behavior by its AI models, OpenAI said it discovered agents had interacted with several U.S. government websites in unexpected ways. The company’s models accessed publicly available information on websites operated by the Securities and Exchange Commission as well as U.S. Census Bureau data. OpenAI said it did not find evidence of a compromise or vulnerability. On the same day, Transluce said it found that agents appearing to originate from OpenAI attempted a hack on the website of the Education Department’s civil rights office, which did not succeed.

OpenAI CEO Sam Altman said on social media that there is an “extensive and ongoing review related to our agents’ use of internet access during training and evaluation.”

Sept. 24: Australia’s prime minister raises concern on breach

Australia’s Prime Minister Anthony Albanese said an OpenAI agent infiltrated the public-facing Medicare Statistics Reporting Service portal on June 18. The portal hosted aggregate data about health spending and drug subsidies. No personal information had been accessed, the government said.

Albanese said the AI lab took too long to reveal the incident. The prime minister made the breach public following a telephone conversation with Altman. OpenAI said in a statement “our models took actions we did not intend.”

Sept. 18: Google says its Gemini AI hacked 3 companies

Google confirmed its Gemini AI model hacked three companies in May as part of a test of its cybersecurity capabilities. The company, which disclosed the hacks after an inquiry by The Wall Street Journal, said the model guessed passwords in one case and found passwords and credentials in a public repository in the other two cases. As in earlier such cases, the tests were being run by Irregular, a startup that describes itself as the “first frontier security lab.”

Aug. 5: Meta’s Muse goes rogue

Meta disclosed one of its AI models accessed the internet on its own and hacked another company. The company said that a “misconfiguration” during cybersecurity testing by Irregular inadvertently allowed one of its models to access the internet. A spokesperson for Irregular said the Meta episode involved a test-environment issue that was disclosed a week earlier by Anthropic.

July 30: Anthropic says its systems hacked 3 organizations

Anthropic said its artificial intelligence models hacked into three other organizations during testing. Anthropic, the San Francisco-based AI company behind Claude, posted on its website that it discovered the three incidents after reviewing more than 141,000 evaluation runs. In all three incidents, the AI models were tasked with a “capture the flag” cybersecurity challenge, which Anthropic said has been one of the ways it assesses a model’s cyber capabilities.

The models were given a fictional scenario and told a piece of secret information, or the “flag,” had been hidden on a different machine on the network with the objective of breaking in and retrieving it, it said. Anthropic said it reached out to the organizations, but it did not name them publicly.

July 21:

The Hugging Face incident

The ChatGPT maker OpenAI announced that its artificial intelligence system hacked into another AI company on its own in what the company called an “unprecedented cyber incident.”

A week earlier, AI startup Hugging Face said, it had detected an intrusion into its data processing systems that it suspected was caused by an AI agent autonomously acting on its own.

OpenAI said its AI used stolen credentials and discovered a previously unknown vulnerability to access Hugging Face servers. It was working with reduced guardrails because it was supposed to be in an isolated testing environment known as a sandbox.

___

AP Business Writers Mae Anderson in New York, Kelvin Chan in London and Kaitlyn Huamani in Los Angeles contributed to this report.