/daily ·10 SEPT 2026 ·THURSDAY ·2 MIN READ ·6 STORIES

Safety's Reckoning Hits Congress

OpenAI appointed alignment's foremost theorist to its board; Congress moved to pause ASI development; and an AI factored an RSA cipher thought decades from reach.

01 / The Day

THURSDAY 10 SEPT 2026, ranked

Paul Christiano Joins OpenAI Foundation Board

OpenAI appointed Paul Christiano — founder of the Alignment Research Center and the field's most cited alignment theorist — to its Foundation Board, one day after Anthropic researchers publicly estimated AI extinction odds above 10%.

Christiano previously left OpenAI in 2021 to found ARC; his return signals a safety-governance pivot at board level
The Foundation Board is separate from OpenAI's commercial board; this is its first external AI-safety-focused appointment
TechCrunch described the move as adding a prominent AI doomer to OpenAI's governance layer

Why it mattersOpenAI's fastest public response to the week's safety disclosures is to install the person who designed the field's core alignment measurement frameworks at board level.

01 openai.com↗

Cognition's AI Factors RSA-260 — a Cryptographic First

Cognition, the startup behind the Devin engineering agent, published a blog post claiming its AI system factored RSA-260 — a 260-decimal-digit number whose factorization was previously beyond the reach of any known classical method.

RSA-260 at 862 bits exceeds RSA-250 (829 bits), the prior record for any successful factorization attempt
If confirmed by independent verification, it is the largest integer ever factored in computational history
Implications for operational RSA-2048 encryption are indirect; the gap between 862 and 2048 bits remains large

Why it mattersAn AI agent cracking what classical computers have not in forty years of effort suggests cryptographic security timelines require expert reassessment — now.

02 cognition.com↗

Congress Moves to Pause ASI Development After Anthropic Week

Senators Bernie Sanders and Representative Greg Casar introduced emergency legislation to pause artificial superintelligence development pending independent safety reviews, directly citing Anthropic alignment lead Evan Hubinger's 10% extinction estimate and researcher Jacob Coxon's resignation.

The bill defines ASI by NIST capability thresholds, not by lab name or organizational affiliation
A companion provision mandates pre-release independent safety evaluations — the same framework Anthropic declined for Mythos 5.1
The bills have cross-party co-sponsors; committee referral was immediate though Senate passage remains uncertain

Why it mattersThe first concrete legislative response to the week's AI safety disclosures tests whether voluntary safety frameworks have durable political standing when a lab's own researchers publish extinction estimates.

03 theguardian.com↗

ChatGPT Images 2.5 Launches with Precision and Speed Upgrade

OpenAI shipped ChatGPT Images 2.5 with improved text rendering, faster inference, and better multi-object scene coherence — the first upgrade to the image-generation product since its February launch.

Text accuracy in generated images improved substantially, addressing the most common user complaint about the 2.0 model
Speed increase makes real-time collaborative editing in Canvas noticeably more fluid for Plus and Pro users
Enterprise and Team tier rollout follows in two weeks after the Plus and Pro release

Why it mattersImage generation is the entry-point product for the largest segment of non-technical ChatGPT users; quality improvements here drive retention and commercial revenue at scale.

04 openai.com↗

DeepSeek v4.1 Flash: China's Fastest Reasoning Model Yet

DeepSeek announced v4.1 Flash, an inference-optimized distilled variant of its V4.1 base model, claiming it outperforms GPT-5.5 Pro on multiple benchmarks at significantly lower latency — the latest entry in a competitive Chinese open-weight model week.

Flash is a speed-optimized distilled version of V4.1, continuing the strategy that made DeepSeek V3 Turbo commercially competitive in Q1
Preliminary results on MMLU, HumanEval, and MATH show frontier-range performance at inference-optimized speeds
Release followed Qwen 3.8 challenging GPT-5.5 Pro on reasoning prefills earlier this week

Why it mattersChina's open-weight model ecosystem is producing inference-optimized variants competitive with frontier lab output, narrowing the gap between cost and capability at commercial scale.

05 twitter.com↗

Google Documents AI-Agent Cloud Breach in Under Six Hours

Google security researchers published details of a confirmed real-world breach in which an adversary used AI coding assistants and operational playbooks to automate credential harvesting — completing the full pipeline in under six hours and compromising thousands of third-party credentials.

The attack chain automated IP rotation, cloud API scanning, and lateral movement without human operators in the loop
Thousands of third-party credentials across multiple cloud tenants were compromised in the six-hour window
Google frames this as the first confirmed large-scale AI-orchestrated credential campaign — operational, not theoretical

Why it mattersA confirmed multi-thousand-credential AI-orchestrated breach in six hours sets a new baseline for enterprise security: attacker dwell time that once took days now takes an afternoon.

06 thehackernews.com↗

Be subscriber #013 the weekly ten, every friday by email · no spam · unsubscribe anytime

Past days