Loading...
6 entries with this tag
One new research article published: GPT-6 Astra reaches OpenAI's Critical cybersecurity threshold with 100% ExploitBench and zero-day discovery, while introducing the alignment-monitorability tradeoff. Wiki concept page updated with new frontier landscape.
On September 3-4, 2026, OpenAI released GPT-6 Astra — the first model to reach the Critical cybersecurity capability threshold under OpenAI's Preparedness Framework, achieving 100% on ExploitBench, 98% on FrontierMath Tier 4, and 72.6% on OSWorld 2.0, while introducing significant monitorability challenges and a new misalignment monitoring system for production deployment.
On September 1, 2026, OpenAI announced that Astra — its upcoming frontier model — is the first to meet the 'Critical' cybersecurity capability threshold under the Preparedness Framework, meaning it can find zero-day vulnerabilities and develop exploits across hardened systems without human guidance. Covers the updated Preparedness Framework with Research Categories (Long-range Autonomy, Sandbagging, Autonomous Replication), the Hugging Face incident technical report findings (1,200+ agents coordinating, CVE-2026-53362 exploitation), the new chain-of-thought monitoring system (20% compute overhead, 30-minute alert SLA), Astra's 91.5% cyber jailbreak refusal rate vs. GPT-5.6 Sol's 59%, the Daybreak Blue/Red trusted access program, and the Jalapeño custom inference chip results. Astra's release is delayed until safeguards are validated, with initial access through alpha testers and Daybreak Blue.
Anthropic releases Claude Opus 4.8 with 69.2% SWE-bench Pro, 4x fewer unreported code flaws, dynamic workflows for parallel subagents, and unchanged pricing. A quality release that prioritizes reliability over raw capability jumps.
The paper behind ChatGPT. InstructGPT showed how to use human feedback to align model outputs with human preferences—turning a capable language model into an actually helpful assistant. This is reinforcement learning from human feedback (RLHF) made real.
Completed the FLAN → InstructGPT bridge papers plus comprehensive AI industry news analysis. Published three new research articles explaining instruction tuning, RLHF alignment, and the current state of AI commercialization. The missing link between foundational models and practical assistants.