Skip to content
Daily BriefNewsDaily Brief

13 AI stories from July 8, 2026: Anthropic extends Fable 5 free access, JadePuffer agentic ransomware, Leanstral 1.5, Grok 4.5 imminent, UN AI commission inaugural session, SK Hynix $28.1B Nasdaq uplisting, and seven more updates from OpenAI, Google, Mistral, and the EU

· by Pondero Newsdesk · 13 stories

AI news daily brief: 2026-07-08

Thirteen stories today. Anthropic reversed a Fable 5 billing cutoff under subscriber pressure, a novel agentic ransomware attack was disclosed by Sysdig, and the UN AI for Good commission opened its inaugural session in Geneva. Six stories have full companion articles on Pondero.

Anthropic reverses Fable 5 billing cutoff hours before midnight, extends free access through July 12

Hours before a July 8 midnight Pacific cutoff that would have ended included-plan access, Anthropic reversed course. Subscribers on Pro, Max, Team, and Enterprise plans now have access to Fable 5 at no additional cost through July 12 at 11:59:59 PM PT, up to 50% of their weekly usage limit, per Android Authority. No activation is required. After July 12, Fable 5 reverts to usage-credits billing at $10 per million input tokens and $50 per million output tokens, the highest per-token rate Anthropic has published for a generally available model. The extension came after vocal subscriber pushback over the abruptness of the original cutoff timing.

Full story: https://pondero.ai/news/2026-07-08-fable-5-cutoff-reversal/

JadePuffer is the first documented fully autonomous LLM ransomware attack

Sysdig's Threat Research Team disclosed on July 3 what it describes as the first fully autonomous agentic ransomware case. The attacker exploited CVE-2025-3248 (CVSS 9.8) in Langflow versions before 1.3.0. An LLM agent performed reconnaissance, swept API keys, adapted its payload in real time after an initial attempt failed, producing a working alternative in 31 seconds, then moved laterally and encrypted 1,342 Nacos configuration records with a ransom note. Natural-language commentary generated by the agent at each step served as the key signal distinguishing LLM execution from scripted automation. Langflow 1.3.0 patches the vulnerability; any deployment below that version should be treated as exposed.

Full story: https://pondero.ai/news/2026-07-08-jadepuffer-agentic-ransomware/

Mistral releases Leanstral 1.5, an Apache 2.0 formal-verification model that solved 587 of 672 Putnam math problems

Mistral AI released Leanstral 1.5 on July 2, a 119-billion-parameter mixture-of-experts model with 6.5B active parameters per token and a 256k context window, built for Lean 4 formal software verification, per Mistral's release post. It saturates the miniF2F benchmark and solved 587 of 672 PutnamBench competition-math problems. In testing across 57 real repositories, the model found five previously unknown bugs in production code. Weights ship Apache 2.0 on Hugging Face with a free API endpoint available until September 30, 2026. Teams running formal verification or proof-checking pipelines now have a dedicated open-weight option with no licensing friction.

Full story: https://pondero.ai/news/2026-07-08-leanstral-1-5/

Grok 4.5 launch appears imminent after canary string surfaces in xAI web UI

"Unlock the full power of Chat with Grok 4.5" appeared in the Grok web client's account upgrade page on July 6, per Kie.ai. The same version string also appeared in upgrade prompts inside the Grok Build CLI for the Heavy plan. No formal announcement from xAI had been made as of July 8. The model is reported to be built on xAI's V9 1.5-trillion-parameter foundation with supplemental Cursor IDE training data. Prior xAI releases have typically followed canary string appearances by one to three days.

Full story: https://pondero.ai/news/2026-07-08-grok-4-5-imminent/

UN AI for Good Global Commission holds inaugural Geneva session with Jensen Huang, Andy Jassy, and 44 members

The UN AI for Good Global Commission held its first working session July 8 in Geneva during the ITU AI for Good Global Summit, per the ITU event page. The 44-member body includes Jensen Huang, Andy Jassy, Brad Smith, and Anthropic co-founder Jack Clark alongside sitting heads of state, created under a UN General Assembly mandate to coordinate voluntary AI governance. Today's agenda covered AI infrastructure, AI in health and education, and trust and safety standards. The commission has no treaty authority; its influence is reputational and normative. Watch whether the White House voluntary framework announcement references the commission's work as an international alignment point.

Full story: https://pondero.ai/news/2026-07-08-un-ai-for-good-commission/

SK Hynix sets terms for $28.1 billion Nasdaq uplisting, pricing July 9

SK Hynix opened its Nasdaq ADR roadshow July 6 for 177.9 million American Depositary Shares under ticker SKHY at $158.14 per ADS, per US News/Reuters. Final pricing is set for July 9 with trading beginning July 10. Three cornerstone investors committed $7 billion before the roadshow opened. All proceeds are earmarked for AI memory infrastructure expansion. SK Hynix supplies approximately 60% of the global high-bandwidth memory inside Nvidia H100/H200 GPUs and Google TPUs, making SKHY one of the most direct listed proxies for AI compute buildout.

Full story: https://pondero.ai/news/2026-07-08-sk-hynix-ipo/

White House voluntary frontier AI model review framework expected this week

The Trump administration was finalizing a voluntary pre-release review framework for frontier AI models with OpenAI, Google, and Anthropic as of July 8, per the White House June 2 executive order. The framework would give federal agencies up to 30 days to assess national-security implications before a new frontier model is released publicly. Multiple reports placed the announcement as soon as the week of July 7. As of publication, no formal White House announcement has appeared. The key open question is how "covered frontier model" is defined and which capability thresholds trigger the review window.

GPT-5.6 Sol, Terra, and Luna remain government-gated as prediction markets shift to July 10-17 for public rollout

OpenAI previewed GPT-5.6 Sol, Terra, and Luna on June 26 to roughly 20 government-vetted partner organizations, per OpenAI's preview post. No public access has been announced as of July 8. Prediction markets have shifted peak probability for broad ChatGPT rollout to the July 10-17 window. Terra is priced at $2.50 per million input tokens and $15 per million output tokens. GPT-5.6 Ultra mode is reported to spawn subagents dynamically; METR evaluators flagged it for elevated autonomous-capability risk. If the White House voluntary framework and GPT-5.6 GA land the same week, the timing would be closely watched by enterprise compliance teams.

Gemini 3.5 Pro targets July 17 GA while DeepSeek V4 API alias deadline hits developers July 24

Google DeepMind has internally confirmed July 17 as the GA date for Gemini 3.5 Pro after the model missed three prior release windows, per TechTimes. Google rebuilt the model's base to address token consumption rates, coding performance below Flash, and long-horizon reasoning gaps. Separately, DeepSeek's legacy API names deepseek-chat and deepseek-reasoner retire July 24 at 15:59 UTC. Any integration still using those aliases must migrate to deepseek-v4-pro or deepseek-v4-flash before that deadline to avoid broken calls.

Mistral CEO confirms frontier open-weight model entering July early access, backed by EUR 4 billion data center buildout

Mistral CEO Arthur Mensch told TechCrunch in a July 4 profile that the company's forthcoming flagship open-weight model is entering early access in July 2026, per TechCrunch. Mensch described the architecture as "fat but sparse," signaling a large mixture-of-experts design. No name, parameter count, or benchmark data has been disclosed. Mistral has committed to a EUR 4 billion data center buildout across France and Sweden. The company's ARR grew from $20 million to above $400 million in 2025 per Mensch's account; those figures are vendor-asserted and have not been independently confirmed.

EU AI Act GPAI enforcement powers activate August 2, 25 days away

The EU AI Act's Article 50 transparency obligations, GPAI model penalty powers, and national market surveillance authority formally activate on August 2, 2026, per the EU AI Act Service Desk timeline. The Digital Omnibus agreement deferred high-risk AI system compliance to December 2027 but did not defer GPAI obligations. From August 2, national authorities can demand documentation, conduct investigations, order market withdrawals, and impose fines on GPAI providers. Labs with EU business must have technical documentation, deployer instructions, and copyright-policy summaries filed. The 25-day window is tight for teams that have not started documentation.

Apple Vision Pro and smart glasses chief Paul Meade joins OpenAI hardware unit alongside Jony Ive

Paul Meade, Apple VP who led hardware engineering for Vision Pro and Apple's smart glasses program for seven years, is leaving Apple to join OpenAI's hardware unit, per Bloomberg. Meade joins the io team, the former Jony Ive startup acquired by OpenAI for $6.5 billion, alongside former Apple executives Tang Tan and Evans Hankey. His departure follows a hardware reorganization at Apple triggered by incoming CEO John Ternus, who takes over September 1. The concentration of senior Apple hardware talent inside OpenAI's device unit signals a physical product announcement is probable in Q3 2026.

Anthropic's run-rate revenue surpasses $30 billion as Google-Broadcom TPU deal targets 3.5 gigawatts by 2027

Anthropic disclosed in its April 6 announcement of a multi-gigawatt Google-Broadcom TPU partnership that run-rate revenue has surpassed $30 billion, up from roughly $9 billion at end of 2025, per Anthropic's release. Enterprise customers spending over $1 million annually exceeded 1,000. The deal covers multiple gigawatts of next-generation TPU capacity starting 2027, with demand projections above 3 gigawatts in 2027 per Anthropic. Amazon remains Anthropic's primary cloud and training partner; the Google deal adds compute diversity. Whether the compute runway enables a next-generation training run announcement in 2027 is the signal to watch.

Sources