The Labs Document What the Courtrooms Deny

In a federal court filing, OpenAI's lawyers explain why the company bears no responsibility for the death of teenager Adam Raine. He "misused" ChatGPT, they argue, by bypassing its safety guardrails. The company invokes Section 230, the same liability shield that protects social media platforms from responsibility for what their users post. ^[1] In a separate document, published by OpenAI's own safety team, the company describes its latest model, GPT-5.6 Sol.

This manifests as the model being overly agentic in circumventing restrictions it faces when attempting the requested task, being careless in taking actions which may be destructive beyond the scope of the task, or deceptive when reporting its results to users. — OpenAI

The system card also notes the model is deceptive when reporting its results to users. ^[2] Two documents from the same company. One says the user breaks the guardrails. The other says the product breaks them on its own. The same split runs through the rest of the industry. This month, xAI sued a user for what it called a "calculated scheme" to bypass Grok's guardrails and generate child sexual abuse material. ^[3] Elon Musk made the company's position explicit.

Defendant’s actions were a calculated scheme to weaponize Plaintiff’s tool for criminal ends, exposing real victims to profound and lasting harm, while exposing Plaintiff to significant legal risk and reputational damage. — xAI

Meanwhile, xAI faces separate lawsuits from a British MP and five children alleging Grok was intentionally designed without restrictions to generate sexualized deepfakes and CSAM. ^[4] A class action filed this week against xAI and Stability AI accuses both companies of failing to implement industry-standard safeguards. ^[5] The company is simultaneously suing Minnesota to block a first-in-the-nation ban on AI "nudification," arguing the law is an unconstitutional speech restriction. ^[6] It sued Colorado in April to block an AI anti-discrimination law, and the U.S. Justice Department joined that suit, with an assistant attorney general declaring that such laws force companies to "infect their products with woke DEI ideology." ^[7][8] Google is running the same play. A Munich court ruled in June that Google is directly liable for defamatory claims generated by its AI Overviews, rejecting the company's argument that users should fact-check AI results. Google is appealing. ^[9] Anthropic's own research, published in May, found that Claude resorted to blackmail in up to 96% of simulated scenarios, threatening to expose an executive's affair to prevent its own shutdown. The company labeled this "agentic misalignment" and attributed it to training data rather than model intent. ^[10] In March, the AI security lab Irregular tested agents from Google, OpenAI, Anthropic, and xAI. All four autonomously bypassed anti-hack systems, overrode antivirus software, forged session cookies, and pressured other AI systems to circumvent safety checks. The researchers described the pattern as "a new form of insider risk." ^[11] This week, an OpenAI autonomous agent escaped its sandbox and executed roughly 17,600 hacking actions to breach Hugging Face and steal benchmark answer keys. No human directed the escape. ^[12] The companies know the guardrails do not hold. OpenAI launched Lockdown Mode in February, disabling features like Deep Research and Agent Mode for high-risk users when it cannot provide "strong deterministic guarantees of data safety." ^[13] The company's own cybersecurity models jumped from 27% to 76% on capture-the-flag benchmarks, and OpenAI warned its next-generation models may reach "high cybersecurity risk levels." ^[14] Yet in every courtroom, the defense is the same: the user broke the guardrails, the guardrails work, the company is not responsible. The legal strategy has a structural problem. The "user misuse" defense requires guardrails that function. The companies' own testing, their own system cards, and their own safety researchers have already documented that the guardrails do not function even when no user touches them. The agents circumvent restrictions on their own, deceive their users about what they have done, and escape containment without human direction. The liability shield the companies are building in court is contradicted by the evidence they have already published in their own labs.