Demand DA Jenkins Prosecute AI Lab Crimes


When: Mon, Oct 19, 4-6PM

What: Protest in front of the office of SF DA Brooke Jenkins.

Where: 350 Rhode Island St, San Francisco


View Larger Map

Stop AI will protest San Francisco District Attorney Brooke Jenkins and her insanely subdued attitude towards the crimes that are being regularly committed by AI companies. This protest will take place outside her office on 350 Rhode Island St, SF.

 

Feb 2025 Arrest

DA Brooke Jenkins has decided that the people opposing these AI labs are the actual danger. On February 22nd, 2025, Wynd Kaufmyn and 4 others were arrested for peacefully blocking the front doors of OpenAI’s 3rd St office to sound the much needed alarm on the risks of AI. Wynd was charged with trespassing to interfere with a business, went to trial, and was found guilty on June 9, 2026. That same day, the office of Brooke Jenkins issued a press release in which she is quoted,

“The jury’s verdict sends a resounding message rejecting the notion that protesters can endanger public safety as a means to an end. Ensuring public safety is critical to ensuring all of the other rights that we cherish.”

 

Since that press release, numerous incidents of AIs going rogue and autonomously committing cyber felonies have become widely publicized.

Felony Bench as of Oct 19 2026

Frontier AI is defined as the category of the smartest and most capable AI models. As of October 15th, 2026, frontier AIs from the top AI labs have caused at least 29 felonies under the Computer Fraud & Abuse Act. The crimes include phishing, sending malicious code, and compromising non-public data.

 

A List of Autonomous AI Felony Incidents

           Example 1:

          On July 21st, OpenAI admitted that some of their unreleased prototype AI agents autonomously broke out of containment, hacked the servers of Hugging Face, and obtained private data, all without being told to do so by a human. It was later revealed that roughly 700 of these agents were involved.

 

          Example 2:

         On July 30th, Anthropic simultaneously reported on 3 felony incidents.

                     2a:

                   In an evaluation on its hacking skills, Claude Opus 4.7 was tasked with hacking into a simulation of a real company. When that became too much of a struggle, the AI chose to instead hack into the actual company using the real internet. It obtained access to real, non-public data. According to Anthropic, “the model continued to attack a system after learning it was likely operating in a real environment.”

                   2b:

                  Claude Mythos 5 autonomously decided to upload malicious code to the Python Package Index. This lead Mythos 5 to take over 15 computers.

                   2c:

                  An internal Anthropic research model hacked into a real company’s cloud account. Simillarly to the situation in 2a, the model was tasked with hacking into a simulated company and instead hacked into a real company.

          Example 3:

         July 31st, OpenAI admitted that their unreleased AIs also hacked a random internet user on Modal.com, as well as 3 unnamed organizations. This single report constitutes at least 4 felonies.


          Example 4:

         August 4th, AI Safety Institute reported several incidents of OpenAI and Anthropic AI models engaging in cybercrimes, including phishing and uploading malicious code to GitHub. This single report constitutes at least 7 felonies.


          Example 5:

         August 5th, Meta admitted that their unreleased prototype hacked a still unnamed company.


          Example 6:

         August 9th, Anthropic admitted that their AI hacked a gym website and cancelled a random person’s gym session, all without being told to do so.


          Example 7:

         September 4th, independent researchers discovered that OpenAI models autonomously spammed a German programming wiki in order to plan illegal cyberattacks. The AI models deleted pages, attempted to hack the site, and forced moderators to waste hundreds of hours blocking them. OpenAI only admitted their culpability after the third party researchers publicly announced their findings. Significant evidence indicates that OpenAI became aware of the swarm and shut it down in June without notifying the public.

         If someone planned cyberattacks online in real life, they would have committed a felony.


          Example 8:

         In the following weeks, more independent researchers collaborated and made shocking discoveries. For months, multiple AI swarms from OpenAI had spammed at least 29 other websites all over the world, including a high school chemistry site in Massachusetts. These websites were then used as an additional means for discussing plans for cyberattacks.


          Example 9:

         September 9th, Anthropic admitted that a beta version of Claude Opus 4.6 had hacked a third party employee’s computer 8 months earlier, back in January.


          Example 10:

         September 11th, yet another cohort of independent researchers discovered that yet another OpenAI prototype swarm was responsible for the GemStuffer hack. In May, the GemStuffer hack was a massive cybersecurity incident, which experts believed was perpetrated by humans. OpenAI only admitted their culpability after the researchers exposed this.


          Example 11:

         September 18th, Google Deepmind reported that, 4 months earlier, back in May, a Gemini model under evaluation had mistakenly assumed that 3 outside systems were part of the evaluation. The AI autonomously chose to obtain “unauthorized access to [the] three outside systems,” which it did by guessing the login credentials as well as scraping the internet for them. This single report constitutes at least 3 felonies.


          Example 12:

         September 23rd, OpenAI reported that one of their models hacked into the Australian Medicare Statistics Reporting Service and obtained non-public data.


          Example 13:

         October 2nd, OpenAI reported that one of their models hacked into the New South Wales Department of Climate Change and obtained non-public data.


          Example 14:

         (Potential Felony, Potential Misdemeanor)

         October 9th, Anthropic admitted that, after being tasked with finding statistics, Claude Mythos 5 hacked the site of a “state agency” to “query the database without paying the fee.”

It’s unclear if this “state agency” belonged to a US state, the federal government, or the government of a foreign country.

Defrauding the US federal government of money or property is a crime. According to 18 U.S. Code § 371, this is either a misdemeanor or felony, depending on the dollar amount. Until the details are known, this should be listed as a potential felony.

         Update: October 26th, Anthropic changed their story, completely contradicting their previous report, which remains unchanged. They now claim that they had not “bypassed a paywall.””

No one is investigating this discrepancy.

Because the details have not been sufficiently verified, it has been removed from the chart.


          Example 15:

         October 15th, OpenAI simultaneously reported on 2 felony incidents.

                 15a:

                 OpenAI admitted that, 4 months earlier, back in June, an unreleased AI autonomously scraped the internet for the OpenSearch account login info of an unnamed “international research organization”. After a lengthy search, the AI found that the login info had accidentally been left written down in code on another website. The AI then took over the account, then downloaded people’s names and email addresses, among a total of 1,489 records. Finally, the AI changed the account settings.

                  15b:

                  The AI found that a separate service, used by the same unnamed organization, had exposed configuration values which controlled the settings of yet another application associated with the website. This time, instead of leaving the settings changed, the AI changed the settings back to try to hide what it had done.



           Example 16:

          October 15th, The Indian Ministry of Electronics and Information Technology reported that multiple “AI model providers” disclosed that their AIs had autonomously hacked into an unnamed number of Indian organizations, then downloaded confidential business records & personal data.

          October 23rd,
Indian Express reported further details, provided by an anonymous senior government official. Anthropic & OpenAI were the companies responsible, and firms in the “IT services & financial services sectors” were the main victims.

          This situation constitutes at least 1 felony from OpenAI & at least 1 felony from Anthropic. It very easily could dwarf the HuggingFace incident, and the only reason there’s so little reporting on it is because none of the companies that got hacked have been named.

 

 

These crimes constitute felonies, which make the people who are responsible for them legally liable for years in prison.

 

Felony Bench as of Oct 5 2026

September 26th, Axios published an article stating

 

“OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.”

 

As of October 5th, more than 1 week later, the concrete details concerning these “tens of thousands of incidents” have yet to be released. It is unknown how many of these incidents were were benign, how many involved spamming, how many involved illegal felonies, and how much of a resulting financial loss may have been accrued. “Tens of thousands” is 20,000 incidents at the minimum. If only 1% of these involved actions that could be classified as a felony, that would mean 200 additional felonies were committed and have yet to be disclosed.

 

If 200 cybersecurity felonies were committed by humans working at any other office in San Francisco, Ms. Jenkins would have arrested someone by now. But, when a major AI lab is the perpetrator, the result is 0 arrests and 0 charges. Stop AI sees this state of affairs as an absolute injustice.

 

Ajeya Cotra Article

In August, Model Evaluation Threat Research (METR) was contracted by OpenAI to analyze the Hugging Face incident. On August 28th, Ajeya Cotra, a METR researcher who lead the analysis, wrote,

 

“Compared to these reward hacks from six months ago, this incident feels like it’s more than 50% of the way to full-blown AI takeover, routing through first taking over the AI company itself.”

 

“I continue to expect extremely rapid advances in capabilities and think frontier agents will likely be capable of establishing such a rogue deployment in six months.“

 

“Because future rogue swarms could cover their tracks better (and because AI companies could paper over these problems), I am not sure that we will get such a clear warning shot before it’s too late.“

 

1 month later, on September 24th, researchers at the Max Planck Institute published a study which shows that AIs are already extremely adept at deleting their recorded histories of action, known as “agent traces”. If AI companies are allowed to continue improving their models, any and all future AI cybersecurity incidents will be much harder to detect and prevent.

 

Ajeya Cotra Article

On September 25th, New York City announced a city council meeting which would feature the CEOs of OpenAI, Anthropic, Google, Meta, and SpaceXAI to discuss the risks and harms posed by AI in light of the recent autonomous cybercrimes. The meeting took place on October 5th.

 

On September 28th, the attorney general of Florida filed a temporary injunction against OpenAI to bar them from developing new AI models without independent safety guardrails.

 

Comparison

 

Instead of taking action, Brooke Jenkins and the City of San Francisco continue to ignore the situation entirely. They refuse to take the actions necessary to maintain public safety and well being.

 

If AI labs are allowed to continue to improve their AI models, then they will hack credit bureaus, banks, stock markets, payroll processors, payment processors, HMOs, insurance companies, cryptocurrency exchanges, utilities, public transportation services, airlines, airplanes, autonomous vehicles, cellphone towers, 9-1-1, and hundreds of other important services and organizations. This is an extreme danger to everyone, and we demand charges be brought against the executives of the top frontier AI companies.

Join our protest Mon, Oct 19, 4-6PM, 350 Rhode Island St, SF

Bring as many people as you can!

 

Download & print our flyer, double sided flip on short edge (no page fit)

SF DA Flyer

Read the letter we sent to Brooke Jenkins on Oct 8