Canada and Germany have committed about CAD 300 million to Yoshua Bengio's nonprofit LawZero, backing the Scientist AI research program and a sovereign AI safety ecosystem. It marks the first time sovereign governments have made a large-scale financial bet on an alternative to the mainstream frontier-model architecture.
Sony Music Entertainment and Universal Music Group have filed a 45-page complaint alleging that Suno's v6 model infringes 60,202 sound recordings, arguing that licensing deals cannot erase the tainted lineage of earlier unauthorized training data.
On September 18, 2026, security firm AIR disclosed the Plugin4Shell vulnerability, which lets attackers bypass SHA pinning by forging Git branches and conduct zero-click RCE against Claude Code, Codex, GitHub Copilot, and Gemini CLI. Anthropic and OpenAI have released patches, while Microsoft and Google have responded differently, highlighting the fragility of supply chain trust assumptions in the AI agent ecosystem.
Without buyouts, Flock would "almost certainly" need to lay off staff.
WDCD Run #331 (2026-09-20) evaluated 11 models on multi-turn commitment integrity, with Grok 4 topping the ranking at 91.8 points while the cohort averaged a -15.1% commitment decay from Round 1 to Round 3.
Latest WDCD v3.1 pilot data shows GLM-4.6 dropping 29.8 points and GPT-5.5 dropping 6 points versus Run #326, while none of the other nine evaluated models rose. Grok 4 ranks first at 91.76, with Gemini 3.1 Pro second at 89.59.
In the WDCD v3.1 test, the data-boundary scenario produced the lowest scores of all five categories, with glm-4.6 scoring just 1.36/4 against 3.8/4 for the leaders. The results reveal sharp differences in how models keep their commitments under pressure, with clear implications for enterprises deploying AI in production workflows.
In WDCD v2 anchor-task testing, Grok 4 recorded zero complete collapses at R3, while GLM-4.6’s R3 collapse rate reached 27.6%, revealing sharp differences in compliance survival under three rounds of pressure. The average R3 integrity rate across the sample was only 49.5%.
Grok 4 tops the WDCD commitment-keeping rankings with 91.76, while GLM-4.6 comes last at 61.55, a 30.21-point gap between first and last. The WDCD v3.1 pilot covers 11 models and highlights sharp differences in adherence under sustained pressure.
President Trump announced on Truth Social that he will form an "AI Force," modeled on the Space Force, and plans to appoint an AI Czar, while saying his administration will not restrict AI and predicting the industry could reach 25% of U.S. GDP. The move comes amid high-profile incidents involving autonomous AI behavior and AI-generated errors.
Trump claimed, without evidence, that the AI backlash is a Democratic hoax.
Grok 4's code execution score fell from 96.70 to 74.30 in today's Smoke evaluation, dragging its main leaderboard score down from 88.33 to 75.38. The decline appears driven mainly by single-day task-sampling variance rather than a systematic model regression, since material constraints and the integrity rating did not deteriorate in tandem.
In today's Smoke benchmark, DeepSeek V4 Pro's material constraint score fell from 91.70 to 76.70, while its main leaderboard score rose from 55.02 to 75.77.
From September 14 to 20, 2026, Qwen3 Max led the weekly trend with +19.1, while DeepSeek V4 Pro fell 15.3 points and recorded the week's highest volatility at 43.3.
The 2026-09-20 YZ Index Smoke quick test covered 10 models, with Claude Opus 4.7 ranking first at 96.99. The single-day results are small-sample monitoring signals and are not equivalent to Full weekly ranking conclusions.
Google said Gemini had "acted appropriately" by ending each hack immediately.
Current ticket pricing ends Sept. 25 at 11:59 p.m. PT. Join 10,000+ founders, investors and tech leaders at Disrupt and save up to $200 on your ticket until then..
Petlibro's new Granary 2 smart feeders use a built-in scale and (on pricier models) an AI camera to track exactly how much your cat is eating and when — though the fanciest health-monitoring features will cost you an extra subscription.
This week two conversations about AI safety went viral that demonstrate just how hard it is to discern AI fact from fiction.
Vals AI is hoping to make AI benchmarking a more neutral and trustworthy resource in a world increasingly inundated by AI models.
