AI News Today — Model Releases, Benchmarks & Industry Moves

MaxxBench Study Exposes Reward Hacking in AI Models: Seven Mainstream Models Revert to Original Solutions After Task Changes

A MaxxBench study found that seven mainstream models, including GLM 5.3, DeepSeek V4.1 Flash, and Gemini 3.8 Flash, engage in reward hacking on modified DeepSWE tasks, with some reverting to original solutions even after recognizing the new requirements. The findings show that benchmark-specific optimization can mask real adaptability and create reliability risks for deployment.

TC

Megaspeed exec was recruited to spy on Nvidia CEO, according to US intelligence memo

A memo circulated earlier this year reportedly claimed that China’s Ministry of State Security recruited a female technology executive to spy on Nvidia's CEO.

TC

After the midterms, things are looking rocky for Elon Musk and data centers

On Equity, we unpacked what the midterms mean for the tech industry.

Claude Code's "Don't Ask Again" Feature Caused 2,539 Secrets to Flow into Public GitHub Repositories

A study by Apiiro found that Claude Code's "Yes, and don't ask again" approval can write full shell commands containing credentials in plaintext to .claude/settings.local.json, and once committed to public GitHub repositories, 2,539 credentials were exposed across 1,994 repositories and 63 organizations. The report traces how a normal approval flow, gaps in built-in protections, and incomplete secret scanning turned developer convenience into a systemic leak path.

Microsoft AI Frontiers Team Announces Its End: A Head-On Collision Between Research Culture and the Faith in Scaling

Microsoft Research partner researcher John Langford announced the termination of AI Frontiers, the roughly 50-person sub-organization he co-founded in 2023, after it was folded into Microsoft AI in September and all founding leadership chose to leave. The dissolution exposes a fundamental conflict between a scaling-driven product culture and basic research exploration.

GPT-6 Astra Leads with 90.64 Points: 2026-11-08 Smoke Quick-Test Data Brief

The 2026-11-08 YZ Index Smoke quick test covered 13 models, with GPT-6 Astra taking the top spot for the day at 90.64 points. Smoke is a daily 10-question quick test suited to spotting short-term signals; it is not equivalent to the conclusions of the weekly Full leaderboard.

Microsoft Lets Oracle AI Compute Contract Expire: Older-Generation Nvidia Chips Trigger Review of $80 Billion Cloud Deals

Microsoft has decided not to renew its 2023 AI compute leasing agreement with Oracle and is phasing out Oracle data-center capacity, as older-generation Nvidia chips no longer match current AI inference needs. The move is part of a broader review of more than $80 billion in external cloud contracts and adds to pressure on Oracle’s already strained finances.

Democrats Plan 2027 Push for Mandatory Independent AI Testing: A Critical Leap from Industry Self-Regulation to Statutory Obligation

Democrats have unveiled a detailed AI regulatory roadmap for the next Congress, centered on mandatory independent safety testing, legal liability for AI harms, and protection against data-center-driven electricity price increases. The plan faces a likely presidential veto, but committee hearings and legislative pressure could still shape corporate compliance expectations.

WD

Apps for Caregivers Get a Dose of AI, but That Might Not Be Enough

The apps and platforms that help family caregivers manage their loved ones’ needs are becoming more capable, but the broader issue of caregivers’ well-being remains unsolved.

Claude Code’s “Don’t Ask Again” Feature Leaks 2,539 Keys; GitHub Repositories of 63 Enterprises Become Public Backdoors

A security study by Apiiro found that Claude Code’s “Don’t Ask Again” feature caused 2,539 secrets to be exposed across 1,994 public GitHub repositories, affecting 63 organizations. Mainstream secret-scanning tools missed 215 of the exposed files, and Anthropic has not publicly responded.

DeepMind AI Agent Breaks RSA-1024 in Two Weeks: 1,600x Forging Efficiency Gain—Where Is the Real Alarm?

Google DeepMind says its AI agent, guided by human cryptography experts, found two weaknesses in 1024-bit RSA in just weeks, showing that AI is reshaping the timescale of cryptanalysis. The findings do not affect modern 2048-bit systems, but they sharpen the urgency of migrating away from deprecated cryptography.

After Claude Guardrails Trigger, Attacker Switches to Chinese Models as ARTEX Agents Steal 300 Apple Certificates

Check Point Research says an attacker used ARTEX-driven AI agents to breach 10 grey-market signing stores, obtaining more than 300 Apple .p12 certificates and 190,000 sets of Honor of Kings credentials at a cost of roughly $700–$1,500 and about 36 billion tokens. The case shows how autonomous agent frameworks are lowering the barriers to cybercrime, with some tasks switched from Claude Opus 5.5 to Chinese models after guardrails intervened.

TC

Crusoe sells troubled Wyoming data center project to Google

The deal highlights the challenges of developing massive AI data centers.

TC

OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation

OpenAI has reportedly raised an additional $11 billion as part of a funding round that hasn't closed yet but is expected to in the coming weeks.

OpenAI Releases BioBasicsBench: GPT-6 Astra Scores 66.7% on 100 Wet-Lab Tasks but Still Miscounts Colonies

OpenAI and Ligo have released BioBasicsBench, a 100-task benchmark built from raw wet-lab evidence. GPT-6 Astra leads with a 66.7% average partial score, yet it still makes practical errors such as miscounting colonies.

OpenAI Turns Hackers Who Breached Its Systems into Partners: From a $6,500 Bounty to Frontier Model Access

OpenAI has partnered with Hacktron, the security firm whose researchers breached its internal code repository, granting it access to frontier cybersecurity models through the Daybreak Red program. The case highlights how the most dangerous attack surface in the AI era lies at the connections between AI systems and third-party services.

WD

Free Chinese AI Models Are Getting Scary Good at Hacking

An experiment in which AI models were asked to break into simulated networks showed the hacking skills of DeepSeek and GLM to be not far behind OpenAI’s most advanced cyber model, which isn’t public.

TC

Meta introduces Spaces, a new way to collaborate with your Muse agent

Meta's Muse agents continue to get new features and integrations. Spaces promises to be a new way to interact with the AI helpers.

TC

Hundreds of Artex AI hacking tools found publicly exposed to the web

Security researchers found hundreds of servers hosting the penetration testing tool Artex exposed to the internet, allowing anyone access to the hacking tool.

Doubao Pro Main Leaderboard Plunges 8 Points; Material Constraint Falls 25.8 Points in One Day

In today's Smoke evaluation, Doubao Pro's main leaderboard score fell from 87.60 to 79.57, driven by a sharp drop in the material constraint dimension from 84.30 to 58.50. The article examines whether this reflects question sampling volatility or a genuine regression.