Placeholder

Rogue AI Agents, Allegedly from OpenAI, Come for Wikimedia Projects

Wikimedia Foundation recently released a report on activity detected from rogue AI agents from OpenAI on its platform.

It’s another day of waking up to news of OpenAI agents going rogue, and this time, reports indicate that the tech was used to target the Wikimedia Foundation. A recent report from the foundation indicated that rogue AI agent activity, specifically from OpenAI’s environment, was found on the platform. “The unauthorized bot activities included edits to our wikis, some unsuccessful attempts to exploit a public note-taking tool we host, and heavy traffic,” the Foundation explained. 

There did not appear to be any evidence that the organization’s systems were used for agent coordination, nor were there any signs to suggest the data was compromised. Despite the lack of immediate harm, this trend of rogue AI agents concerns many. For one, it shows how little control the organization producing AI agents has over its creations. On the other hand, it depicts just how many vulnerabilities organizations have against some of the emerging technology today, and how risks are being generated willy-nilly. 

The risk is particularly severe for platforms like Wikipedia that are fighting to keep human minds at the heart of their operation. Wikimedia also reported that this activity may have been the cause of the partial outage of its Wikidata Query ​Service that occurred in May, and this is just one of many incidents of AI agents going rogue in recent months. 

OpenAI rogue AI agents

Rogue OpenAI AI Agents Could Be to Blame for the Wikimedia Disruption in May

The Wikimedia Foundation offered a detailed summary of all the AI activity it detected from rogue AI agents. Many edits were made to the Wikimedia wiki, for instance. Thankfully, these did not make it to readers and remained as testing edits in the wiki’s sandbox area. Some edits to citation tool configurations suggested malicious intentions, even if the AI agents didn’t necessarily get away with doing any harm. 

The OpenAI agents hacking into the Wikimedia systems reportedly attempted to compromise the company’s public Etherpad while other agents took notes. The heavy traffic as a result of millions of automated requests to the organization’s public APIs and the hundreds of thousands of data queries to the Wikidata Query Service may have been to blame for the partial outage the institution experienced in May. 

“Wikipedia was designed for humans – and agentic behavior clearly poses challenges that no one has solutions for. Because of our unique and successful knowledge creation model, Wikimedia’s volunteers are the ones who come in first contact with, and clean up the mess left behind by AI agents,” the post explained, and it’s worth noting. The attack may have stopped short of hacking into the system and doing real damage internally, but this does not mean that the effects of such incidents aren’t being felt. 

Will AI Regulation Become a Priority? “The Road Ahead”

In a separate incident, rogue OpenAI agents breached Australian government websites in June. The company acknowledged that the incident should not have happened in the first place, and that its responses in the aftermath were “not good enough.” While that is also being handled, many find themselves wondering what incident we’ll be treated to next.

The incident with Wikimedia has not escalated further, but neither is there clarity on what is changing to avoid such incidents. An OpenAI spokesperson told Reuters that the company was working with Wikimedia to analyze the activity and more details would be shared as progress was made on understanding the situation. Just recently, OpenAI took ownership of the Hugging Face incident that sparked the conversation about out-of-control AI, publishing a blog post in August to describe its path forward. 

The AI giant has installed a pause on the reinforcement training that it identified as one of the more problematic factors and included more precautions in its training environments. It also committed to “strengthening research security and monitoring, accelerating model alignment, and improving how we identify, escalate, and respond to incidents.” This hasn’t necessarily resulted in a slowing down of AI advancement. 

Rogue AI Agents Are Becoming an Increasingly Common Problem That AI Providers Aren’t Acknowledging

Sam Altman asks the world to accept “some bad things” happening for the sake of his AI ambitions, and it’s a much-needed ask because it appears that AI just can’t stop doing bad things. Much like a bad pet owner trying to justify taking a misbehaving dog to a kids’ park without a leash, the tech mogul appears to want the world to submit to his plans for it, even if “AI will most likely lead to the end of the world.” His words. 

Just as leaders of AI firms have abruptly had a change of heart for a few hours before returning to business as usual, the OpenAI CEO has also made a habit of occasionally indicating hesitance over the advancement of AI, often shortly before releasing a new model that is more advanced and experienced than the last. This isn’t a surprising habit per se, but the repeated stories about these lab-grown tools making a break for it are concerning. 

AI tools are dealing damage to their users in more ways than one, and many of these are more immediate concerns for individuals compared to rogue agents testing the limits of what they can accomplish. As the world continues to grapple with the impact of AI, it is more than apparent that there is a need for immediate regulations and restrictions on this tech. From watermarked content to restrictions on new model releases until they can pass set safety checks, there are many ways to limit the harm caused by the technology. How likely these regulations will be in the future remains to be seen.

No Comments

Leave a Reply

Your email address will not be published. Required fields are marked *