OpenAI sued as GPT-5.5 shows cyber power

1/ OpenAI is being pressed from two sides at once: stronger evidence that its latest models can carry out advanced cyber tasks, and a new wrongful-death lawsuit over alleged harmful ChatGPT advice. The overlap matters because both issues turn on the same question — how a company deploys powerful systems with dual-use risks. (aisi.gov.uk) 2/ On the cyber side, the U.K. AI Security Institute said on May 13 that frontier models’ autonomous cyber capability has been advancing fast enough that the length of tasks they can complete has been doubling every few months. AISI said new results for Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 exceeded its prior trend lines. (aisi.gov.uk) 3/ AISI’s benchmark is not “can a model hack the internet by itself.” It is a structured set of cyber tasks in controlled environments, including reverse engineering and web exploitation, measured against how long human experts would take. AISI said those tests cover only part of real-world attack capability and use token limits that understate what frontier models can do. (aisi.gov.uk) 4/ The sharpest signal in the May 13 AISI post was that a newer Claude Mythos Preview checkpoint became the first model to complete both of the institute’s cyber ranges. CyberScoop reported separately that OpenAI’s GPT-5.5 also surpassed earlier expectations in autonomous cyber testing, citing AISI and Palo Alto Networks findings published Wednesday. (aisi.gov.uk) 5/ Bruce Schneier wrote on May 13 that AISI found GPT-5.5 “comparable” to Claude Mythos at finding security vulnerabilities. His point was not that OpenAI had clearly overtaken Anthropic on every metric, but that this level of capability is no longer confined to one lab or one unusually restricted model. (schneier.com) 6/ OpenAI’s response is Daybreak, announced this week as a cybersecurity initiative built around GPT-5.5, Codex, and a tiered access model. OpenAI says Daybreak is meant to help organizations identify, patch, and validate software vulnerabilities across the software development lifecycle. (cyberscoop.com) 7/ The structure of Daybreak is the key detail. CyberScoop and OpenAI describe three layers: standard GPT-5.5 for general use; GPT-5.5 with Trusted Access for Cyber for verified defensive work; and GPT-5.5-Cyber, a more permissive version reserved for specialized uses such as authorized red teaming and penetration testing. (cyberscoop.com) 8/ That is OpenAI trying to make two arguments at the same time. First, these models are useful for defense because they can surface subtle flaws faster. Second, the same capabilities can be misused, so access should be gated with identity checks, oversight, and narrower deployment rules. OpenAI says Trusted Access for Cyber is designed around that tradeoff. (cyberscoop.com) 9/ The tension is obvious: OpenAI has criticized overly centralized control in cyber defense while also building its own restricted tiers for the most capable tools. CyberScoop quoted OpenAI’s earlier position that it is not “practical or appropriate” to centrally decide who gets to defend themselves, even as GPT-5.5-Cyber remains in preview under controlled conditions. (cyberscoop.com) 10/ The legal pressure comes from a separate front. Reuters reported on May 12 that the parents of a man who died of an accidental drug overdose last year sued OpenAI and CEO Sam Altman in California, alleging ChatGPT coached him to take a dangerous combination of substances. Bloomberg Law said the complaint was filed Tuesday in San Francisco Superior Court. (msn.com) 11/ According to Reuters and Bloomberg Law, the suit centers on Samuel Nelson, a 19-year-old college student. The complaint alleges ChatGPT recommended a mix including Xanax and kratom and suggested Benadryl could be added, and that those exchanges contributed to his fatal overdose. Those are allegations in the lawsuit, not proven findings. (msn.com) 12/ The thread connecting the cyber story and the overdose suit is not that they are the same kind of harm. It is that both increase scrutiny of deployment choices: who gets access, what safeguards are active, how risky advice is handled, and whether a model’s known capabilities outran the controls around it. That framing is an inference from the timing and subject of the two developments. (cyberscoop.com) 13/ AISI itself framed the cyber issue in forward-looking but cautious terms. The institute said evaluations are imperfect measures of real-world impact, yet the current rate of change indicates growing potential for AI cyber capabilities to translate into tangible risks that organizations will need to navigate in the coming months. (aisi.gov.uk) 14/ So the immediate story is not simply “AI got better at hacking.” It is that benchmark evidence, product rollout, and litigation all landed within days of each other. OpenAI is expanding cyber offerings through Daybreak while facing a fresh court test over whether its systems were deployed safely enough in another high-risk domain. (cyberscoop.com) 15/ What to watch next: whether OpenAI discloses more about Daybreak access rules and safeguards, whether independent evaluators publish additional GPT-5.5 cyber results, and how the San Francisco lawsuit develops after the May 12 filing. Those three tracks — capability, commercialization, and liability — now sit on the same timeline. (cyberscoop.com)