Last year, according to sources who chose to remain unidentified, The New York Times reported that hackers managed to access the online forum utilized by OpenAI staff for secure internal conversations. From the forum postings, hackers extracted specifics about the design of the company’s AI technologies, however, they did not breach OpenAI’s actual AI storage and production systems.
During an all-hands meeting in April 2023, OpenAI executives shared details about the breach with all employees and also notified the board of directors. Since there was no theft of information regarding customers or partners, the incident was not disclosed to the public.
Sources indicated that law enforcement was not informed because executives believed the hacker was not affiliated with a foreign government, deeming the breach as not posing a threat to national security.
A representative from OpenAI, in an email to TechRepublic, stated: “As we communicated to our employees and board last year, we identified and resolved the underlying issue and are continuing to enhance our security measures.”
How did select OpenAI staff respond to this data breach?
Reports from the NYT suggested that news of the forum’s security compromise raised concerns among other OpenAI employees, who feared it revealed a vulnerability that could be exploited by government-backed hackers in the future. The possibility of OpenAI’s advanced technology falling into the wrong hands and being misused for threatening reasons was a significant worry.
SEE: OpenAI’s GPT-4 Can Autonomously Exploit 87% of One-Day Vulnerabilities, Study Finds
Additionally, the handling of the incident by executives led some employees to question whether OpenAI was adequately safeguarding its proprietary technology from external threats. An ex-technical manager, Leopold Aschenbrenner, claimed to have been terminated after expressing these concerns to the board of directors on a podcast with Dwarkesh Patel.
OpenAI refuted these allegations in a statement to The New York Times, including disputing Aschenbrenner’s interpretation of their security measures.
Latest Updates on OpenAI Security, Including Insights on the ChatGPT macOS App
Besides the forum breach, recent events suggest that security might not be a primary focus at OpenAI. Data engineer Pedro José Pereira Vieito recently revealed that the new ChatGPT macOS app was storing chat data in plain text, potentially allowing unauthorized access to the data if a Mac fell into malicious hands. Upon being informed of this vulnerability by The Verge, OpenAI promptly issued an update encrypting the chats as a preventive measure.
An OpenAI spokesperson informed TechRepublic via email: “We acknowledge this issue and have released an updated version of the application that encrypts these conversations. We are committed to delivering a user-friendly experience while upholding our stringent security standards amidst our technological advancements.”
SEE: Millions of Apple Applications Were Vulnerable to CocoaPods Supply Chain Attack
In May 2024, OpenAI disclosed that it had thwarted five covert influence operations originating from Russia, China, Iran, and Israel that aimed to exploit its models for deceptive activities. These activities involved generating comments and articles, creating fictitious names and bios for social media accounts, and translating texts.
During the same period, a Safety and Security Committee was established by the company to design the protocols and safeguards necessary for the development of their avant-garde models.
What Does the OpenAI Forums Breach Suggest About Potential AI Security Incidents?
Dr. Ilia Kolochenko, Partner and Cybersecurity Practice Lead at Platt Law LLP, expressed his belief that this breach of OpenAI forums might be just one of numerous security incidents involving AI technologies. He communicated to TechRepublic via email: “The global AI competition has transformed into a national security concern for many nations, leading state-sponsored hacker groups and mercenaries to aggressively target AI vendors, from promising startups to tech behemoths like Google or OpenAI.”
Dr. Kolochenko stated that hackers aim for valuable AI intellectual property, such as robust language models, training data sources, technical research, and commercial data. He mentioned the potential implementation of backdoors to manipulate or disrupt operations, similar to recent attacks on critical infrastructure in Western countries.
He added: “All users of GenAI vendors in enterprises must exercise caution and prudence when sharing or allowing access to their proprietary data for LLM training, as this data — ranging from attorney-client privileged information and trade secrets of major industrial or pharmaceutical firms to classified military data — is also in the sights of AI-hungry cybercriminals who are prepared to escalate their attacks.”
Are There Effective Approaches to Mitigating Security Risks in AI Development?
There is no straightforward solution to completely warding off security threats from foreign adversaries during the creation of new AI technologies. OpenAI cannot discriminate based on employees’ nationalities, nor does it wish to limit its talent pool by restricting hiring to specific regions.
Moreover, it is complex to predict the misuse of AI systems before malevolent intentions surface. A study by Anthropic revealed that LLMs were only slightly more beneficial to malicious actors for generating or designing biological weapons than standard web access. An analysis from OpenAI reached a similar conclusion.
On the contrary, some experts contend that while not an imminent danger, AI algorithms could pose risks as they advance further. In November 2023, representatives from 28 nations endorsed the Bletchley Declaration, advocating for global collaboration to address the challenges posed by AI. The document highlighted the potential for severe, even catastrophic, harm — either deliberate or inadvertent — stemming from the most potent capabilities of these AI models.
About Author
