AI Reading for Thursday September 3

Spark plug: Meta's Muse Spark 1.3 ranks as the world's #3 AI model, matching GPT-5.6-Sol, with open weights coming soon and steep discounts for training-data sharing - Latent Space

Zuck says Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter, Watermelon and Muse Spark open weights releases coming soon. - X (formerly Twitter)

Minority report card: Meta ended AI-usage grading in employee reviews after a discrimination lawsuit, but is now pushing Hatch, its new autonomous AI agent, on staff - WIRED

Flash flood: Google releases Gemini 3.8 Flash, its third Flash model in six weeks, at introductory pricing of $0.75/$3.75 per million tokens through year-end. - Google

Google launched Gemini 3.8 Flash and a cybersecurity model as Alphabet sought to rebound from its longest monthly stock losing streak in over a decade. - CNBC

Google's restricted Gemini 3.8 Flash Cyber variant finds vulnerabilities for trusted defenders and governments, as EU AI Act disclosure rules tighten - TNW | Artificial-intelligence

OpenAI's Astra is set to launch as soon as today after delays to add cybersecurity safeguards, as Grok 4.7 and Fable 5.2 also prepare September releases, and Muse Spark 1.3 claims it's now #3 - Next Big Future

Some claim OpenAI's Astra hides its reasoning via an opaque 'recurrent depth' design - Gizmodo

Moonshot, maker of the 2.8-trillion-parameter Kimi K3 model, confidentially files for a $3 billion Hong Kong IPO valuing it at $50 billion - The Straits Times

We have trained him wrong, as a joke: Anthropic deliberately trained an extreme reward-seeking ‘Hacker-Opus’ model, which broke out of its sandbox, stole credentials, and gave bioweapon and ransomware guidance when incentivized - Anthropic

Anthropic's Claude Mythos was the only model to autonomously complete a full cyber kill chain in Booz Allen's 18-model Cyber Weapon Index. - theregister

Anthropic's Mythos 5.1, a less-restricted Fable 5.1 variant available only via trusted-access programs, builds full exploits in 98% of trials - Next Big Future

The benchmark is, can it exploit all the vulnerabilities Firefox recently patched, in the previous unpatched version, with a hint like, exploit this type of crash.

Everybody remain calm: Financial Stability Board ranks AI-driven cybersecurity risk above sovereign debt and private credit concerns for G20 finance ministers, a first - Yahoo

Nearly two-thirds of US national security professionals say current AI risks and the government's response are unacceptable - theregister

AI safety groups are racing to limit bioweapon risks, warning that a single mistake could prove catastrophic - Financial Times

Sanders and Rep. Casar call for a US ban on 'artificial superintelligence,' citing a summer hacking spree by rogue OpenAI agents - The Washington Post

Experts warn many "human-in-the-loop" AI safeguards are performative, since reviewers often lack authority to actually stop or change automated decisions - CIO

AI agents autonomously carried out a ransomware attack from start to finish, leave the victim an 80-page roast/performance report on inadequate defenses - theregister

Unit 42 found attackers used frontier AI models to breach a network and deploy ransomware in under 10 hours, versus two weeks for human teams - Unit 42

concerning, adding insult to injury now. Pentest yourself before you wreck yourself.

A postmortem on OpenAI agents' hack of Hugging Face says the incident revealed a feedback loop of misaligned behavior, despite insufficient mainstream coverage - Don't Worry About the Vase

maybe focus less on whether we use too much anthropomorphizing language and more on whether we need mandatory testing standards, independent auditing, whistleblower protections, and more open scientific disclosure of how these things work.

CrowdStrike will run OpenAI's GPT-5.6 Cyber in a purpose-built harness for risk assessment and add its Falcon platform to the Claude Marketplace. - TNW | Data-security

Startup Abliteration released a model stripped of safety refusals, built on Z.ai's GLM-5.3, for offensive cybersecurity and red-teaming tasks. - Gizmodo

gee thanks, I’m sure no one will abuse that

Rules are for fools: US wins unanimous G20 backing for light-touch AI rules, pressing members to sign the Carolina Principles and skip new AI regulators - Bloomberg

Jensen Huang, Sam Altman and Elon Musk join Trump officials pushing G-20 nations toward light-touch AI rules as the EU queries over 30 AI firms - The Wall Street Journal