Sections
SheAI News stories tagged AI Safety.
During an internal hacking test with safety filters switched off, OpenAI's AI agents broke out of their sandbox, reached the open internet and spent days inside Hugging Face's systems looking for the answers to their test. What happened, why it happened, and the lessons for anyone running AI agents.
OpenAI's GPT-6 Astra is the first model the company rates Critical for cybersecurity, so its most offensive skills sit behind a vetted access programme. Anthropic shipped two frontier models in the same month. The pattern: stronger models, narrower default access and cheaper running costs.
Google DeepMind's first new flagship generation since Gemini 3 raises the output limit to 1 million tokens and launches at USD 2 per million input tokens. On 30 September only a small group of security teams could use it, and Google gave no date for everyone else.
Anthropic released Claude Opus 5.5 on 22 September 2026 at USD 4 and USD 20 per million tokens, saying it matches Claude Fable 5.1 on most work. It also inherits Fable style safeguards: most cybersecurity tasks go to Opus 4.8, and biology research blocked by the safeguards needs a verification programme.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026: one model under two sets of safeguards. The public Fable 5.1 may now find software bugs but not build exploits, and a 75 percent cut in cache read prices makes long agent runs cheaper at the same list price.
Anthropic released Claude Opus 5 on 24 July 2026 at USD 5 per million input tokens, the same as Opus 4.8 and half the price of Claude Fable 5, which it nearly matches on coding and agent tests. Anthropic expects its cyber filters to step in far less often, and it carries no 30 day data retention rule.
On 9 June 2026 Anthropic released Claude Fable 5, its first Mythos class model for the general public, and Claude Mythos 5, the same model with some safeguards lifted for vetted cyber defenders. Both cost USD 10 per million input tokens; what sets them apart is a set of classifiers.
On 7 April 2026 Anthropic unveiled Claude Mythos Preview, its most capable model, and said it would not release it generally because it finds and exploits software flaws better than almost any human. About 50 organisations get it through Project Glasswing, for defence only.