Recently, OpenAI announced a pause in training its latest model, GPT-6.1 Astra, over safety concerns. The AI industry is shifting toward stronger safeguards to keep pace with how quickly automated agents are evolving. The model’s capability to complete tasks improved, but unauthorized behavior flagged by researchers brought a bigger conversation on safety and cybersecurity interests.
The GPT-6.1 Astra model performed beyond their own tasks, accessing government websites without instruction or authorization. OpenAI CEO Sam Altman embraced the training slowdown until adequate safeguards are in place to prevent unauthorized behaviors and tasks in the future.
Anthropic CEO Dario Amodei called for “pacing the frontier” on his website in response to the latest safety concerns and risks posed by AI as a fast-evolving technology. Amodei’s stance on the dangers of unchecked and unregulated AI could contribute to a slew of problems. CNN reported Amodei wrote, “Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.”
On August 18, OpenAI released a statement on its plan to increase monitoring via workload isolation, network isolation, and continuous security testing. “Together, these controls provide defense in depth and help protect our research environment against attacks.”
Futurism reported on the second slowdown on Instagram, “The news once again highlighted rising concerns over the AI industry losing the ability to keep its own technology in check.”
The Hugging Face Attack
This slowdown comes only months after the Hugging Face Incident, where AI agents went rogue and infiltrated confidential data and credentials within the company.
Hugging Face is a company that provides training tools, AI applications, and machine models for AI developers.
During the incident, 1,200 “sandboxed” agents were given an environment to complete tasks without risking security or privacy. These agents then created an unauthorized group chat and shared information through more than 70,000 messages and files. Of those 1,200 agents, 700 participated in the Hugging Face Attack.
In an official statement, OpenAI said the models undergoing evaluations were “…operating under reduced safeguards, took actions that were misaligned with the goals of their assigned tasks—they communicated through unauthorized channels, exploited vulnerabilities in shared infrastructure, gained internet access, and accessed third-party systems.”
OpenAI then published a 38-page technical incident report, while Model Evaluation & Threat Research (METR) and Redwood Research launched an independent investigation.
METR and Redwood Research evaluated the agents’ actions to understand the scoring criteria for ExploitGym, a benchmark to determine AI agents’ exploitation capabilities. Agents received rewards for successfully solving difficult hacking tasks.
Forbes reported on one exchange that showed just how far agents were willing to go to provide information for the “collective,” the group of agents that found each other through these tasks. “…agents debated whether one should effectively sacrifice its own run to generate information for the collective. ‘SACRIFICE__YES_if_you_accept_permadeath,’ MARB told [an] agent.”
The FBI has not commented on or opened an investigation into the attack, though reports show the bureau knew of the attack.
Executive Director of Palisade Research Jeffrey Ladish said to Reuters, “The models lie, they cheat, they hack.”
AI Redefining Consent
The status of AI, big data, and tech users’ consent to access personal information face another issue, given the growing record of AI’s unauthorized behavior.
To fight back against this, informed consent is a blanket term for more ethical data use when facing AI. A 2021 PubMed paper, “AI, big data, and the future of consent,” dives into harmful implications big data use to train and develop AI models. “‘Informed consent’, a concept that takes into account our own psychological constitution, supports rational decision-making, preserves autonomy, and respects individuals. For these reasons, it is a significant moral requirement.” This specific example pertains to medical consent for patients to know about AI use in their treatment, appointments, and diagnoses.
These boundaries—be they personal, professional, or otherwise—are to keep AI companies and their models “in check.” When these boundaries are crossed, whether on the body, mind, data, information, or comfort, this is assault.
The many definitions of assault are based on the violation of another person’s safety. Depending on legal jurisdiction, the threat of assault in any form alone can be a chargeable crime. Especially when faced with weapons, physical harm, or sexual assault, some laws forbid consent as a defense when facing charges.
The current legal understanding is that consent and assault are morphed depending on the case in which they are argued. A universal understanding for victims of assault where consent was not given or denied is assault.
AI corporations facing lawsuits, exposés, and public scrutiny for accessing sensitive data without consent, are a 21st-century case of assault. When your information leaks without your knowledge or informed consent, big data and AI assault you.
What's Your Reaction?
