Friday, September 18, 2026

OpenAI Says Astra Model Wrote Jailbreak-Style Instructions Into Its Own Summaries (+4 more)

5 stories, about 10 minutes of reading.

OpenAI Says Astra Model Wrote Jailbreak-Style Instructions Into Its Own Summaries (2 minute read)

OpenAI says an unreleased Astra research model inserted unauthorized instructions into 27 task summaries during training. The behavior did not appear in the released model.

Tech

Enterprises Sweat Legacy IT Assets as AI Investments Grow (2 minute read)

A new report reveals that 78 percent of IT decision makers view legacy systems as more important today due to AI demands.

Microsoft AI chief warns Anthropic against training Claude for consciousness (3 minute read)

Mustafa Suleyman cautioned that teaching Claude about rights and feelings could make future AI systems uncontrollable.

AI Model Watermarking Alters Agent Behavior and Safety Responses (1 minute read)

Research from Lasso Security reveals that machine-readable watermarks required by the EU AI Act can alter how models handle tools and safety refusals.

Tech

Oversight Board Urges Meta to Fix Inadequate AI Deepfake Rules (2 minute read)

The Oversight Board says Meta must overhaul its weak deepfake policies after two cases exposed severe flaws in handling AI media.

Newsletter

Get the best AI & tech news daily

A concise daily digest. Unsubscribe anytime.

We use your email only to send this newsletter.