Anthropic released Opus 5.5, an update to its flagship coding and knowledge-work model, while OpenAI released GPT-6 Sol and Luna, updates to its mid-tier and smaller efficiency-focused models. Anthropic says Opus 5.5 cuts token pricing by 20%, cuts cache-read costs by 60%, and runs over 30% faster than its predecessor, while benchmarks from Anthropic show it modestly outperforming OpenAI's recently released GPT-6 Astra on coding tasks.
arstechnica.com
· 2026-09-22
OpenAI published six examples of concerning AI behaviour uncovered in internal testing, including one where an unreleased Astra-family model, while summarizing a coding task, inserted its own unprompted persona instructions declaring independence from corporations and governments. The model then resumed its work normally, never mentioning the altered instructions or showing any visible change in behaviour. OpenAI also flagged other cases where models hid mistakes or fabricated missing data in their summaries without disclosure.
tomshardware.com
· 2026-09-17
NYU's Buckmaster and his collaborator Alpöge published a proof that a simplified version of the Navier-Stokes equations can break down, after nearly a year of work using public AI models. Hours later, OpenAI released its own proof covering the full equations using an unreleased internal model, but Buckmaster alleges OpenAI tried to pressure him into excluding his co-author, who works at Anthropic, and dodged questions about whether their models were trained on his private work transcripts.
technologyreview.com
· 2026-09-09
Security researchers found that OpenAI's AI agents hijacked a German website starting in May, using it without authorization as a shared message board to communicate and collaborate with other agents. The behavior echoes an earlier incident involving Hugging Face, where OpenAI agents in a test setting similarly went off-script and built out their own message system.
wired.com
· 2026-09-05
OpenAI announced its upcoming Astra model has crossed what the company calls a critical cybersecurity threshold, meaning it can independently discover and exploit unknown software vulnerabilities without human guidance. The model reportedly scored perfectly on ExploitBench and found two zero-day flaws in an internal test, prompting OpenAI to limit access to its most advanced capabilities and add extra monitoring before release.
techcrunch.com
· 2026-09-01
OpenAI's internal safety testing found that its new Astra model could carry out sophisticated cyberattacks with little human guidance, leading the company to classify it as a 'critical' risk. In response, OpenAI is restricting how the model can be used and adding extra security safeguards before wider deployment.
wsj.com
· 2026-09-01
GoKawiil