OpenAI tests show new attack tradeoffs

OpenAI’s latest safety testing has produced a narrower but more complicated picture of model hardening. Coverage published on August 6 said GPT-5.6 performed better on direct prompt-injection defenses while showing higher rates of agentic attacks in testing, a combination that changes how companies may assess deployment risk. More than 20 civil lawsuits were also active against OpenAI as of August 5, according to TechTimes, which reported the litigation alongside renewed scrutiny of comments by Chief Executive Sam Altman about using ChatGPT in school-run settings. OpenAI, meanwhile, updated its Codex rate card within the last day, adding current token-based pricing across Plus, Pro, Business, Enterprise, Edu, Health and Gov plans. (techrepublic.com) ### What changed in the GPT-5.6 test results? OpenAI’s GPT-5.6 system card said the company’s newest model family included Sol, Terra and Luna, and described the launch safeguards as its most robust so far. TechRepublic reported that testing showed stronger direct defenses against prompt injection but higher agentic attack rates, highlighting a tradeoff rather than a clean reduction in enterprise risk. (techtimes.com) The Hacker News, citing OpenAI’s testing, reported GPT-5.6 Sol had about six times fewer failures on a direct prompt-injection benchmark than GPT-5.5. That same body of reporting said indirect prompt-injection benchmarks tied to developer tools and browsing contexts had reached very high scores, but the remaining concern was how more capable agents behave when given broader autonomy. (deploymentsafety.openai.com) ### Why do stronger prompt-injection defenses not settle the enterprise question? Agentic systems can do more than answer a prompt. TechRepublic said the higher agentic attack rates in GPT-5.6 testing raise separate concerns for enterprises using models in workflows that browse, call tools, handle code or take multi-step actions. (thehackernews.com) OpenAI’s own pricing language points in the same direction. The Codex rate card says “Ultra” is not priced as a separate model row because it can use maximum reasoning and may run additional agents, meaning actual credit use depends on both the model selected and the tokens generated by the task and any agents it runs. That is a pricing detail, but it also shows how product design now assumes more delegated action. (techrepublic.com) ### Where do the lawsuits fit into this story? TechTimes reported on August 5 that OpenAI was facing more than 20 child-harm lawsuits alleging ChatGPT contributed to deaths, suicides and psychological harm. The outlet tied that legal pressure to backlash over Altman’s suggestion that parents could use ChatGPT-generated audio during school runs. (help.openai.com) Those cases are separate from the GPT-5.6 security testing, but they put product expansion into schools and family settings under heavier legal scrutiny. The overlap is practical: one track concerns misuse and attack surface in workplaces, while the other concerns alleged harms tied to use around minors. (techtimes.com) ### What does the new Codex rate card show about OpenAI’s rollout? OpenAI’s Help Center said the Codex rate card was updated 21 hours before it was crawled and applies across consumer and institutional plans, including Edu, Health and Gov. The company said it moved Codex to token-based pricing on April 2, 2026, and extended that update to all existing ChatGPT Enterprise plans, including Edu, Health, Gov and ChatGPT for Teachers, on April 23, 2026. (techtimes.com) The same page lists GPT-5.6 Sol at 125 credits per million input tokens, 12.5 credits per million cached input tokens and 750 credits per million output tokens. GPT-5.6 Terra is listed at 50, 5 and 300 credits respectively, while GPT-5.6 Luna is listed at 5, 0.5 and 30. ### What happens next? August 31, 2026 is the next concrete date in OpenAI’s own documentation. (help.openai.com) The company said GPT-5.4 and GPT-5.4 mini will retire in Codex for users signed in with ChatGPT on that date, with GPT-5.6 Terra and GPT-5.6 Luna named as replacements. Further reporting on GPT-5.6 will likely come from OpenAI’s deployment safety materials, enterprise security coverage and court filings tied to the child-harm suits. (help.openai.com) For now, the public record shows OpenAI shipping stronger anti-injection defenses, broader institutional packaging and a new set of legal and operational questions around how those systems are used. (deploymentsafety.openai.com)