OpenAI says it has expanded safety testing around its upcoming model Astra as it "cannot rule out" critical cyber capabilities, potentially delaying launch
THE SO WHAT
Labeling Astra as potentially having “critical” cyber capabilities — and expanding testing enough to risk a delay — moves offensive security from hypothetical to operational concern. If you deploy advanced models, assume regulators and customers will start asking how you prevent dual-use abuse, not just hallucinations.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIPalantir’s stock stages best week since 2024, showing it’s no longer an AI loser
Public markets are starting to reward actual AI revenue, not just AI narrative — Palantir’s rebound is tied directly to “booming demand” for deployed AI solutions. If you’re selling into similar buyers, expect procurement to benchmark you against vendors that can show production usage and spend, not pilots.
Applied AIThe next age of LLMs? Dev gets a small LLM running at 10 tokens a second locally on a $10 microcontroller
A 28.9M-parameter model running near 10 tok/s on a sub-$10 microcontroller — with most weights staying in flash — pushes “good enough” language intelligence into the bill-of-materials noise. Hardware teams building devices, sensors, and appliances should be scoping where on-device LLMs can replace cloud calls and unlock offline, low-latency features.
Applied AIOpenAI pumps the brakes on new Astra model over cybersecurity concerns
Pausing internal work on Astra because internal evals flagged “powerful cybersecurity abilities” is a line in the sand — offensive capability is now a gating factor for model release. If you’re building on frontier models, budget time for capability downgrades, stricter access tiers, and slower rollouts as security standards harden.
Applied AIWatching Roku’s AI channel is like eating from a trough
Roku’s AI FAST experiment — an endless stream of auto-generated content — shows what happens when AI fills airtime without taste or curation. Media operators should treat AI as a tool for programming and personalization, not a replacement for editorial judgment, or risk burning audience trust quickly.