AI/ML Security & Trends
The dominant AI/ML security story is a wave of serious infrastructure breaches: Fireworks AI confirmed attackers reached customer datasets, incomplete model weights, and billing data via a compromised third-party cloud provider (hitting Notion, Cursor, and Harvey downstream), while NEAR AI disclosed that an RCE in the SGLang inference engine let an attacker read confidential-mode requests from inside a trusted execution environment — a stark reminder that TEEs don't protect against vulnerable code running inside them. Separately, Jensen Huang abruptly resigned as NVIDIA CEO amid an affair scandal involving an alleged Chinese intelligence operative, shaking the industry's most important hardware supplier.
Fireworks AI breach widens to expose model weights, datasets, billing data Breaches & Incidents
Fireworks AI's November 9 update revealed that attackers who breached a third-party cloud provider reached its R&D systems, exposing resource metadata, billing/invoice records, and — for a subset of customers — incomplete model weights and training datasets. This expands an earlier October disclosure about stolen API secrets from evaluation jobs. Notion confirmed Fireworks' breach exposed Notion AI chat transcripts in a small number of workspaces, a case study in third-party AI-infra supply-chain risk.
Read at The Hacker News →SGLang RCE breach at NEAR AI exposed confidential-mode inference requests Breaches & Incidents
NEAR AI disclosed on November 9 that an attacker exploited a remote-code-execution flaw in SGLang, the open-source inference engine it runs inside a trusted execution environment for 'confidential mode' requests. Because the vulnerability was in software running inside the TEE, the enclave provided no protection — the attacker could view in-flight requests, including confidential ones, and later pivoted to a deprecated system to grab ~100 users' names/emails, history from 18 legacy agents, and two leaked API keys. A direct lesson for anyone relying on confidential computing as an ML security boundary.
Read at The Hacker News →Jensen Huang resigns as NVIDIA CEO amid alleged spy affair scandal Industry & Trends
NVIDIA announced on November 12 that Jensen Huang resigned as President, CEO, and board member effective immediately, following reports of two affairs — one reportedly involving a woman suspected of being a Chinese intelligence operative, another with an NVIDIA executive. CFO Colette Kress becomes interim CEO; shares fell over 4%. Huang has since been summoned to testify before Congress. NVIDIA says it will keep its 260,000-GPU supply commitment to South Korea despite the leadership change — a major shock to the company sitting at the center of the global AI compute supply chain.
Read at NVIDIA →Nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceX to halt frontier AI Industry & Trends
The Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI filed suit in San Francisco Superior Court on November 13, invoking California unfair-business-practices law to seek an injunction against continued frontier-model development. The complaint cites the labs' own executives' stated risk estimates, including Dario Amodei's prior one-in-four estimate of AI going 'really, really badly,' arguing the industry is self-regulating development of a potentially catastrophic product.
Read at Reuters via Yahoo →Anthropic sets IPO price range at $70–$78, valuing it up to ~$1.02T Industry & Trends
Anthropic's amended S-1, filed November 10, proposes a $70–$78 per-share range for 825 million Class A shares (plus a 123.75M-share underwriter option), implying a basic valuation of roughly $916B–$1.02T and a fully diluted figure near $1.4T. Pricing is expected November 18 with Nasdaq trading beginning November 19 under 'ANT' — a landmark moment for an AI safety-focused lab going public at trillion-dollar scale.
Read at Yahoo Finance →AWS patches SSRF flaw in MCP Gateway Registry exposing internal services AI Security & Safety
AWS disclosed on November 11 a server-side request forgery vulnerability in its open-source mcp-gateway-registry, affecting versions 1.0.3 through 1.34.0 and fixed in 1.35.0. An authenticated attacker could abuse the flaw to pivot into internal network services behind the gateway. The bug adds to a growing list of MCP-infrastructure vulnerabilities as agentic tool-calling gateways become a new attack surface in production AI deployments.
Read at Cybersecurity News →IBM discloses two MCP Gateway auth/credential vulnerabilities in ContextForge AI Security & Safety
IBM published security bulletins on November 12 for two flaws in its ContextForge MCP Gateway: CVE-2026-110279, an improper-authentication bug (CWE-287) allowing credential override, and CVE-2026-110278, an information-exposure issue (CWE-294) caused by missing per-gateway configuration checks during tool invocation and resource reads. Both affect versions 1.0.0 through 1.1.0 and are rated medium severity — another sign MCP gateway implementations are an active vulnerability class.
Read at IBM →JailBound: FOL-guided framework evaluates jailbreak boundaries across 46 LLMs AI Security & Safety
Researchers published JailBound, a unified jailbreak-evaluation framework combining automated benchmark construction with first-order-logic-guided dual-branch search to map LLM 'behavioral boundaries' rather than single pass/fail jailbreak outcomes. The paper, accepted to NeurIPS 2026, evaluates 46 models across 13 model families under one experimental protocol — useful for red-teamers wanting a standardized way to compare jailbreak robustness across frontier and open models.
Read at arXiv →Alibaba launches Qwen4-Max preview with 1M-token context Model & Product Releases
Alibaba opened preview access to Qwen4-Max on November 11 across QwenCloud, Qoder, the Qwen app and QwenStudio, offering a 1-million-token context window and a 393,216-token output budget. The preview is time-boxed, retiring after November 25, and is priced from $3/month for new users — part of a broader Qwen4 lineup (Max, Flash, Plus, 27B) Alibaba is rolling out to compete with Western frontier models on long-context and agentic coding tasks.
Read at OrcaRouter →Google DeepMind ships Gemini Decision 1, a decision-only scoring model Model & Product Releases
Google DeepMind announced Gemini Decision 1 on November 11: a model that takes one prompt plus up to 1,000 candidate answers and scores every option in a single pass, generating no text — just probabilities and a confidence score in 150–250ms. Priced at $0.10 per million input tokens with no output billing, it undercuts Microsoft's rival Decision-1 by roughly 2.4x and follows OpenAI's Decisions API (GA Oct 22) and Amazon's Strands Decider 2B — a new product category for high-throughput classification/ranking workloads.
Read at CellCog →NVIDIA contributes OpenShell agent-sandboxing runtime to CNCF with Red Hat Tools & Frameworks
At KubeCon + CloudNativeCon NA, NVIDIA announced it is contributing OpenShell — a secure runtime that isolates AI agents in sandboxed execution environments and applies deterministic policy controls over tool use, inference access, and workload execution — to the Cloud Native Computing Foundation under open governance. Red Hat is integrating it into OpenShift alongside masked credentials and per-agent identity, explicitly framed as a response to rising prompt-injection and agent-hijacking risk in enterprise deployments.
Read at SiliconANGLE →OpenAI cuts off Cursor's model access after SpaceX's Anysphere acquisition Industry & Trends
OpenAI notified Cursor-maker Anysphere that it was winding down OpenAI model access inside Cursor effective November 12, following SpaceX's acquisition of Anysphere. OpenAI cited concerns that Musk-affiliated entities have a history of violating platform terms of service, pointing to Twitter's post-acquisition contract breaches. Cursor's CEO said OpenAI models were only about 5% of platform traffic; Anthropic, Google, Moonshot, Z.ai, Grok and Cursor's own Composer remain available.
Read at InfoWorld →Rox raises $100M Series B at $1.3B valuation for AI sales agents Industry & Trends
Sequoia-backed Rox closed a $100 million Series B on November 11 at a $1.3 billion valuation, co-led by General Catalyst and March Capital with Prosperity7 joining as a new investor. Rox builds AI agents that run enterprise sales workflows (prospecting, pipeline, deal management) for customers including MongoDB, Confluent, Ramp and Puma — one of the larger agentic-AI funding rounds of the week.
Read at Fortune →BrainwashedLLM automates discovery of chain-of-thought jailbreaks AI Security & Safety
A new arXiv paper, BrainwashedLLM, introduces a weak-to-strong adversarial framework that automates discovery of jailbreaks exploiting the chain-of-thought reasoning traces of Large Reasoning Models, arguing that the visible reasoning process itself — not just final outputs — is a powerful and underexplored attack vector for reasoning-capable models.
Read at arXiv →AttentionSwitch proposes jailbreak defense via sparse attention-head control AI Security & Safety
AttentionSwitch, published November 12 on arXiv, is described as the first jailbreak defense that jointly learns sparse attention-head selection and scaling within a frozen LLM, intervening on specific attention heads rather than fine-tuning weights. The authors report it outperforms existing defenses across several popular models and jailbreak attack types, with particularly strong gains in low-training-data regimes.
Read at arXiv →Norwest raises $4B fund to deepen AI bets Industry & Trends
Norwest, the venture firm backed solely by Wells Fargo, closed a $4 billion fund on November 12 — its eighteenth fund in 65 years and its largest to date — explicitly earmarked for increased AI investment, underscoring how much institutional capital keeps flowing into the sector even amid rising infrastructure-cost and compute-spend scrutiny elsewhere in the market.
Read at TechCrunch →US voters pass 40 of 41 ballot measures restricting AI data centers Industry & Trends
Reporting on November 11 detailed that US voters nationwide approved 40 of 41 local ballot measures aimed at reining in AI data-center construction (water use, power demand, zoning), coinciding with Democrats retaking the House and Senate in early-November midterms — a combination likely to reshape federal and local AI infrastructure policy heading into 2027.
Read at Consortium News →EU's AI regulatory sandbox rules take effect December 2, offer no liability shield Industry & Trends
A new EU regulation establishing AI regulatory sandboxes takes effect December 2, 2026, per reporting on November 12. Crucially, participation in a sandbox does not grant companies any compliance clearance or liability shield under the AI Act — testing results don't certify conformity, keeping full regulatory exposure in place even for firms using the sandbox to validate systems before deployment.
Read at Tech Times →