Today’s top line: The FBI warns open-source AI hacking tools can breach networks in under three hours; Nvidia’s leadership crisis intensifies as allies form a “war room” to lobby for Jensen Huang’s return.
FBI says Artex AI tool used to breach networks in under three hours
Summary: The FBI warned that hackers are using Artex, an open-source autonomous AI penetration-testing platform, to carry out attacks that can escalate from reconnaissance to full compromise “in under three hours.” Artex-linked intrusions have been tied to breaches of South Korean financial firms that exposed at least 68,000 people and to the ConnectWise compromise that gave access to Microsoft 365 environments for at least 2,250 customers.
Digest: Autonomous AI tools let attackers scan and exploit obvious flaws far faster than humans can respond, raising urgency for basic hygiene like multi-factor authentication and credential management – and for defenders to shift from reactive monitoring to proactive hardening. Businesses should prioritise blocking exposed credentials, MFA, and rapid patching.
Source: TechCrunch
Allies build a “war room” to lobby for Jensen Huang’s return to Nvidia
Summary: Allies of former Nvidia CEO Jensen Huang are assembling a “war room” of Silicon Valley executives to campaign for his reinstatement while Nvidia’s board runs an independent probe into his conduct. The push follows Huang’s resignation this week and has drawn public support from leaders including Elon Musk and Tim Sweeney, even as Treasury Secretary Scott Bessent called for a “dramatic shake-up” at the company.
Digest: The public campaign highlights how central Huang’s leadership is perceived across tech and how high-profile governance crises can ripple into policy scrutiny and investor pressure; watch Nvidia’s board actions, the interim CEO Colette Kress’s guidance to the company, and any regulatory or customer response related to compliance in Asia.
Source: Ars Technica
OpenAI begins embedded assessments with external evaluators
Summary: OpenAI has started a pilot of embedded assessments by bringing in outside firms Apollo Research, METR, Transluce and Goodfire to work within the company. The programme will independently evaluate model alignment, safeguards and monitoring systems.
Digest: Embedding third-party evaluators aims to catch risky behaviours earlier in development and deployment, which could raise industry expectations for continuous, independent oversight of frontier models; organisations building or buying AI should watch how findings and disclosures evolve.
Source: OpenAI
OpenAI releases CSAM image classifier and ChatGPT warnings
Summary: OpenAI rolled out a new image classifier for child sexual abuse material (CSAM), ChatGPT warning screens and temporary account blocks, developer guidance, and law enforcement tools. The classifier flagged 48% fewer images while finding the same number of abusive images in early tests, and images it flags were twice as likely to be abusive; ChatGPT will show alerts and impose a 24-hour suspension for attempts to create or access sexualised content about children.
Digest: These features aim to reduce reviewer workload and deter misuse of generative tools for child exploitation, while giving partners free API access to the classifier; businesses embedding generative image or chat features should review developer guidelines and reporting flows to meet safety and legal obligations.
Source: TechCrunch
Hugging Face switches routing to OpenRouter to run open-weight models
Summary: Hugging Face moved its Inference Providers to OpenRouter, which routes requests across providers by live price, speed and reliability and retries on failures. OpenRouter serves more than 800 trillion tokens a month to over 10 million users; Hugging Face says open-weight models rose from 26.8% to 71.7% of tokens served between October 2025 and September 2026, and open-weight inference costs are roughly four times cheaper than closed models.
Digest: The change makes it easier for businesses to run cheaper open-weight models at scale while preserving existing APIs and billing; teams should test latency, cache behaviour and model quality trade-offs when switching providers or prioritising cost.
Source: Hugging Face
Audit finds hallucination detectors fail on small, quantized on-device LLMs
Summary: A practitioner-facing audit tested four hallucination detectors on small, 4-bit quantized models used on-device and found detector rankings do not transfer across models – Spearman rho was -0.80 and Kendall tau-b -0.667. A simple lexical baseline performed similarly to sampling-based methods, and standard label pipelines produced near-chance AUROC (0.46-0.59) on one benchmark while the same detectors scored 0.69-0.89 AUROC on another.
Digest: Detector performance on large research models does not predict performance on the smaller models actually deployed on devices; product teams should validate detectors on their exact model, quantisation and workload rather than rely on published rankings.
Source: arXiv
LoRAWeaver proposes training-free adapter routing by prototype matching
Summary: The LoRAWeaver paper presents a training-free adapter routing method that selects parameter-efficient LoRA adapters by matching adapter-conditioned query representations to task prototypes using cosine similarity. Experiments on LLaMA-2-7B and 13B across a 48-task FLAN-v2 benchmark show the method outperforms other training-free baselines and approaches a trained retrieval encoder without retraining when new adapters are added.
Digest: The approach reduces the operational cost and complexity of managing many task adapters, making adapter-based model updates more practical for dynamic production systems; engineering teams using LoRA pools should consider prototype-based routing to avoid retraining routers.
Source: arXiv
Quick hits
- Sam Altman says industry will be “fine” without Jensen Huang – OpenAI CEO Sam Altman spoke at Semafor’s event, expressed empathy for Huang and said he didn’t know enough to judge whether Huang should return (TechCrunch).
- OpenAI expands ChatGPT for Academic Researchers to 200,000 users – OpenAI will give free access to frontier models to 200,000 researchers worldwide (OpenAI).
- Google adds agentic Ask Maps features for holiday planning – Ask Maps can now build itineraries, find appointments and call local businesses to book slots on Android and iOS (Google).
- LVMH Media Division partners with OpenAI for journalism tools – A strategic partnership will use ChatGPT to support LVMH’s publications and accessibility (OpenAI).