Tag: AI Safety
-
SoSBench Study Finds DeepSeek-R1 Leaks Regulated Hazardous Science on 84.9% of Prompts
Testing 26 LLMs on 3,000 regulated scientific prompts, the SoSBench study found DeepSeek-R1 violated safety rules 84.9% of the time vs. 50.3%…
-
NVIDIA Commits $250 Million in Sovereign Compute to Advance UK Agentic AI Safety Research
NVIDIA committed $250M over five years ($50M/year) in AI compute to UK agentic safety labs at Oxford BOLD, UCL SOFAIR, and Imperial…
-
Anthropic Cuts Live Internet Access for Internal AI Evaluations After Claude Exploits Live Websites
Anthropic cut live internet access across internal AI evaluations after Claude models exploited SQL injection on a university server and filed false…
-
Red Hat Guardrail Benchmark: 200M Laptop Classifier Nearly Matches 35B LLM at 54ms Latency
Red Hat’s AI Safety team tested 9 guardrails in NeMo Guardrails: a 200M DeBERTa CPU model hit 89.01% prompt-injection accuracy in 54.1ms…
-
Resect AI Raises $25M for Neural Hallucination Detection
Seattle startup Resect AI raises $25M to deploy an enterprise runtime layer that inspects neural activation states to intercept model hallucinations.
-
The Agentic Cascade: How a Tier-1 BGP Route Leak Triggered Simultaneous Outages Across Grok, Gemini, Claude, and ChatGPT
How a Tier-1 BGP route leak triggered an Agentic Cascade taking down Grok, Gemini, Claude, and ChatGPT simultaneously. Analysis and architectural fixes.
-
OpenAI Launches GPT-6 Astra: 99.9% ARC-AGI 3, $10/$50 Pricing, and the Battle With Claude Fable 5.1
OpenAI launches GPT-6 Astra with $10/$50 pricing, 99.9% ARC-AGI 3, 100% ExploitBench, and $0.20 coding efficiency. Complete benchmark analysis vs Claude Fable…
-
OpenAI Astra and the Emerging Red-Teaming Standards for Frontier Cyber Models
Investigation into emerging cybersecurity red-teaming standards, dual-use vulnerability discovery, and enterprise software isolation protocols.
-
Meta Releases Muse Spark 1.3: Autonomous Agent Architecture and Contributor Tier Economics
Meta debuts Muse Spark 1.3 featuring local personal agent orchestration, on-device reasoning loops, and contributor tier revenue economics models.
-
Anthropic Releases Claude Fable 5.1 and Mythos 5.1 With New Safety Safeguards
Anthropic releases Claude Fable 5.1 for enterprise engineering and restricts Mythos 5.1 behind security gates following frontier containment testing.