FAR.AI

Sep 22, 2026 On the sidelines of UNGA 81 in New York, FAR.AI convened UN officials, AI developers, and counterterrorism experts to address radicalization and CBRN risks from LLMs.

Sep 18, 2026 We evaluated Qoder CLI against an adaptive project-memory attack and found it successfully jailbroke the model on 27 out of 30 malicious cyber requests.

Sep 17, 2026 Anthropic’s Mythos 5 recently grabbed headlines for trying to talk an open-source repo maintainer into merging malicious code. In this paper, we systematically study the broader threat of how AI persuasion could undermine human control, particularly at frontier AI labs.

Jul 29, 2026 FAR.AI's new AI Security Leaderboard tested the safeguards of four frontier models under identical conditions: Claude Fable 5 and GPT-5.6 Sol held, while Grok 4.5 and Gemini 3.1 Pro each broke for under $300. It launches with the Minimal Standard for Safeguards, a shared security floor built entirely from defenses peer developers already run in production.

Jul 28, 2026 FAR.AI's Seoul Alignment Workshop showed a field trying to turn AI safety into an engineering discipline, and discovering how far its measurements still have to go.

Jun 30, 2026 Scaling the SOLiD lie-detector oversight protocol across Llama and Qwen shows that undetected deception drops as models grow and trusted labelers can be removed, provided the detector's training data covers the fine-tuning distribution.