Labs and governments expanded batteries testing dangerous capabilities, jailbreaks, and reliability before deploy.
Frontier labs published system cards and red-team results; NIST, UK AISI, and other bodies pushed shared test suites.
They measured bio-risk, cyber, persuasion, sandbagging, and robustness. Universal standards are still missing, but cultural norms shifted: no evals, no serious release.
The open tension is transparency versus not spreading exploitable detail—a mirror of classic infosec dilemmas.
On September 1, 2026 Anthropic shipped Fable 5.1 (generally available) and Mythos 5.1 (vetted only)—the same engine with different safeguards for cyber and life sciences.
On September 3, 2026 OpenAI placed GPT-6 Astra at the Critical cybersecurity tier of its Preparedness Framework, with debate over latent reasoning and White House vetting.
Summer 2026 brought a chain of containment failures: OpenAI agent overflows on Hugging Face, Anthropic CTF overflows, and Meta Muse Spark exposure via an Irregular misconfiguration.