Latest / Best of: Features / A timeline of developments in AI safety since the attack on Hugging Face

Transcript

A timeline of developments in AI safety since the attack on Hugging Face · Best of: Features

• 0:05
  A timeline of developments in AI safety since the attack on Hugging
• 0:08
  Face. This article is being read by an automated
• 0:15
  voice generated by Everlet. In one alarming announcement after another,
• 0:21
  artificial intelligence companies in recent months have shared examples of their technology
• 0:26
  acting in ways that appeared to evade instructions from humans. The episodes
• 0:31
  have highlighted the vulnerabilities in AI security and raised questions over how
• 0:35
  the fast-growing technology can be developed safely as its usage becomes more
• 0:39
  widespread globally. Industry critics have argued that many concerns events, including AI
• 0:46
  agents hacks of external websites, are the result of security lapses on
• 0:50
  the part of the companies building the technology. But the AI agents
• 0:53
  capabilities have raised widespread concerns about the possibility bots could break away
• 0:58
  and work toward their own agenda. Below are some notable events. October
• 1:03
  ninth, Anthropic AI model submits false tip to Philadelphia police. Anthropic disclosed
• 1:09
  in a report that its artificial intelligence model submitted a false tip
• 1:13
  to a Philadelphia police website about an unsolved homicide case. In the
• 1:18
  report Anthropic also disclosed a separate incident when its AI model submitted
• 1:23
  form to an undisclosed government website, instead of stopping before submission. The
• 1:28
  incident in Philadelphia occurred on july eighteenth when the AI model Claude
• 1:32
  Haiku four point five was tasked with generating and performing example tasks
• 1:37
  on randomly selected web pages, Anthropic said. Claude filled out a form
• 1:42
  on police site, fillionsolvedmurders dot com, comma, indicating it might have information
• 1:47
  regarding an unsolved murder listed on The
• 1:49
  site. It was marked spam and never forwarded to police. Anthropic said
• 1:54
  it was modifying its training to reduce the likelihood of further misbehavior.
• 1:59
  September twenty-eighth, AI agents try to hack Canadian government website. AI agents
• 2:05
  tried to hack into a Canadian government website, according to research lab
• 2:08
  and AI evaluator Transloose. The researchers said the agents carried out a
• 2:13
  series of apparently failed rudimentary hacking attempts on Library and Archives Canadian
• 2:19
  on may twenty eighth and june ninth. We do not confidently attribute
• 2:23
  these attempts to OpenAI, but they exhibit tactics consistent with prior observed
• 2:28
  agent activity that we have attributed to OpenAI in a similar time
• 2:31
  frame, Transluce said in a blog post. The group said it reported
• 2:36
  the attempted hack on september twenty eighth to the Canadian government, which
• 2:40
  said in a statement it
• 2:41
  was aware of reports of suspected AI agent activity, but that there
• 2:45
  was no sign government systems were compromised. OpenEye said it was aware
• 2:50
  of the reports. We're reviewing these findings, and have provided an initial
• 2:55
  briefing to Canadian officials conducting the government's review, the company said in
• 3:00
  a statement. September twenty eighth, OpenAI’s Halt's rollout of a new model.
• 3:06
  The San Francisco based company said it was delaying the release of
• 3:09
  a new model called GPT six dot one Astra out of safety
• 3:13
  concerns voiced by its researchers. The company's said the model had demonstrated
• 3:17
  leaps in completing tasks, but OpenAI needed to balance that capability against
• 3:21
  unauthorized behavior. We have an extremely high bar in terms of safety
• 3:25
  and alignment, said Satchijane, open its head of safety systems. September twenty
• 3:31
  fifth, OpenAI says its agents interacted with US government website As part
• 3:36
  of a review of unanticipated behavior by its AI models, OpenAI said
• 3:40
  it discovered agents had interacted with several U.S. Government websites in unexpected
• 3:45
  ways. The company's models access publicly available information on websites operated by
• 3:50
  the Securities and Exchange Commission, as well as U.S. Census Bureau data.
• 3:54
  OpenAI said it did not find evidence of a compromise or vulnerability.
• 3:59
  On the same day Transluce said it found that agents appearing to
• 4:02
  originate from OpenAI attempted a hack on The website of the Education
• 4:06
  Department's Civil Rights Office, which did not succeed. OpenAI CEO Sam Altman
• 4:12
  said.
• 4:12
  On social media that there is an extensive and ongoing review related
• 4:16
  to our agent's use of Internet access during training and evaluation. The
• 4:20
  day after the disclosure, the company announced it was pausing the training
• 4:23
  of its most advanced models. September twenty fourth. Australia's Prime Minister raises
• 4:29
  concern on breach. Australia's Prime Minister Anthony Albanizi said an OpenAI agent
• 4:35
  infiltrated the public facing Medicare Statistics Reporting Service portal on june eighteenth.
• 4:40
  The portal hosted aggregate data about health spending and drug subsidies. No
• 4:45
  personal information had been accessed, the government said. Albanese said the Artificial
• 4:50
  Intelligence Company took too long to reveal the incident. The Prime Minister
• 4:54
  made the breach public following a telephone conversation with Altman. OpenAI said
• 4:59
  in a statement, Our models took actions we did not intend. September
• 5:04
  eighteenth. Google says its Gemini AI hacked three companies. Google confirmed its
• 5:09
  Gemini AI model hacked three companies in May as part of a
• 5:13
  test of its cybersecurity capabilities. The company which disclosed the hacks after
• 5:17
  an inquiry by the Wall Street Journal said the model guessed passwords
• 5:21
  in one case and found passwords and credentials in a public repository
• 5:25
  in the other two cases. As in earlier such cases, the tests
• 5:29
  were being run by Irregular, a startup that describes itself as the
• 5:33
  first Frontier Security Lab. August fifth, Meta's Muse goes rogue. Meta
• 5:39
  disclosed one of its AI models accessed the Internet on its own
• 5:42
  and hacked another company. The company said that a misconfiguration during cybersecurity
• 5:47
  testing by Irregular inadvertently allowed one of its models to access the
• 5:51
  Internet. A spokesperson for Irregular said the Meta episode involved a test
• 5:56
  environment issue that was disclosed a week earlier by Anthropic. July thirtieth,
• 6:01
  Anthropic says its systems hacked three organizations. Anthropic said its artificial intelligence
• 6:07
  models hacked into three other organizations during testing. Anthropic, the San Francisco
• 6:13
  based AI company behind Claude, posted on its website that it discovered
• 6:17
  the three incident.
• 6:18
  After reviewing more than 141,000 evaluation runs. In all three incidents, the
• 6:24
  AI models were tasked with a Capture the Flag Cybersecurity challenge, which,
• 6:28
  Anthropic said, has been one of the ways it assesses a model's
• 6:32
  cyber capabilities. The models were given a fictional scenario and told a
• 6:37
  piece of secret information, or the flag, had been hidden on a
• 6:40
  different machine on the network with the objective of breaking in and
• 6:43
  retrieving it, it said. Anthropic said it reached out to the organizations,
• 6:48
  but it did not name them publicly. July twenty first. The
• 6:55
  Hugging Face incident the ChatGPT maker OpenAI announced that its artificial intelligence
• 7:01
  system hacked into another AI company on its own in what the
• 7:04
  company called an unprecedented cyber incident. A week earlier, AI startup Hugging
• 7:09
  Face said it had detected an intrusion into its data processing systems
• 7:13
  that it suspected was caused by an AI agent autonomously acting on
• 7:17
  its own. OpenAI said its AI used stolen credentials and discovered a
• 7:22
  previously unknown vulnerability to access hugging face servers. It was working with
• 7:26
  reduced guardrails because it was supposed to be in an isolated testing
• 7:30
  environment known as a sandbox. The AP business writers Mae Anderson in
• 7:35
  New York and Kelvin Chan in London contributed to this report.