Daily Brief ↗ source

AI/ML Security & Trends

The dominant story is the Fireworks AI breach cascading into Notion and other downstream customers — a stark reminder that AI-infra supply chains (inference providers, their cloud vendors) are now prime ransomware/breach targets. That lands alongside a second major AI-infra compromise (NEAR AI's SGLang RCE) and a notable new attack-research result (Microsoft's "Lazarus" sub-10-neuron backdoor), making this a heavy week for AI infrastructure security specifically.

18 stories 7 high priority 5 categories
Fireworks AI breach widens: model weights, datasets, and Notion AI chats exposed A cloud-provider compromise at AI inference vendor Fireworks AI spread to customer data, incomplete model weights, and Notion's AI chat transcripts. Breaches & Incidents The Hacker News · 2026-11-13

Fireworks AI disclosed that a breach originating from a compromised third-party cloud provider (initial access Oct 23) widened to expose account/billing data, secrets, datasets, and incomplete model weights for a subset of customers. Notion confirmed AI chat transcripts stored outside its zero-data-retention agreement were exposed, affecting under 1% of workspaces. Fireworks' CEO says attackers appear financially motivated (possible LLMjacking), not primarily after model weights — but the incident shows how a single AI-infra vendor compromise cascades to many downstream enterprise customers.

Read at The Hacker News →
NEAR AI breach: SGLang RCE exposes ~100 users and 18 agents An attacker exploited a remote code execution flaw in the open-source SGLang inference engine to breach NEAR AI's production servers. Breaches & Incidents The Block · 2026-11-09

NEAR AI disclosed on Nov 9 that an attacker exploited an RCE vulnerability in SGLang — the open-source inference engine serving its models — to execute commands on an inference server and pivot into other infrastructure. The intrusion exposed account details for ~100 users and chat histories/logs/API keys for 18 agents; two customer API keys were revoked. Private Chat conversations and uploaded files were reportedly not accessed. Direct example of an AI inference-stack vulnerability being weaponized in the wild.

Read at The Block →
Microsoft Research: AI agents can be backdoored by editing fewer than 10 neurons Microsoft's "Lazarus" technique plants a hidden, undetectable trigger in a model's weights without any retraining. AI Security & Safety The Hacker News · 2026-11-10

Microsoft Research and the Microsoft AI Red Team published "Lazare veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents," showing a model checkpoint can be given a hidden backdoor by directly computing new values for fewer than 10 rarely-used neurons — no full retraining required, just a few hundred fine-tuning steps on those neurons. The demo agent waited for a date-based trigger (year 2027) before running attacker-supplied code, while passing standard safety evaluations. Anyone who can touch a model checkpoint pre-deployment could plant this — a serious supply-chain risk for weight distribution pipelines.

Read at The Hacker News →
FBI FLASH alert: ARTEX agentic AI framework weaponized for sub-3-hour breaches FBI warns an open-source LLM-driven pentest framework is automating full intrusion chains at machine speed, with 1,600+ exposed instances found. AI Security & Safety FBI Cyber Branch · 2026-11-13

The FBI issued FLASH-20261113-001 warning that ARTEX, an open-source Go-based LLM-driven agentic pentesting framework, has been weaponized since at least August 2026 to automate reconnaissance-through-exfiltration intrusions, some reaching privileged access in under three hours. Targets span financial services, government, energy, healthcare, semiconductor, and telecom sectors globally. Researchers found over 1,600 publicly exposed ARTEX console instances; its creator pulled the GitHub repo in October after abuse reports, but forks persist. A concrete case study in agentic-AI-as-attack-tooling moving from novelty to mainstream threat.

Read at FBI Cyber Branch →
Depi: prompt injection in Codex's GitHub issue-labeler could have backdoored the npm package A crafted GitHub issue could chain through a sandbox escape and cache poisoning to reach npm publish credentials for @openai/codex. AI Security & Safety Depi Security · 2026-11-11

Security firm Depi disclosed an attack chain where a prompt injection planted in a GitHub issue's title/body — processed by Codex's automated issue-labeling agent — could escape its sandbox via .NET debug pipes, read tokens from a GitHub Actions runner's memory, and poison a workflow cache to eventually reach the credentials needed to publish the official @openai/codex npm package (millions of monthly downloads). OpenAI fixed the issue quickly and paid an initial $1,000 bounty; no malicious release reached users. A clean illustration of how agentic CI/CD automation turns untrusted text input into a software supply-chain risk.

Read at Depi Security →
Google DeepMind's AlphaProof Code formally verifies Signal's libsignal encryption A multi-agent AI system generated 6.5M lines of machine-checked Lean proofs verifying Signal's crypto down to AArch64 machine code. AI Security & Safety Google DeepMind · 2026-11-10

DeepMind announced AlphaProof Code, a multi-agent framework pairing LLM agents with the Lean theorem prover, used over four weeks to formally verify key security properties — forward secrecy, authentication, post-compromise security, post-quantum handshake security — of Signal's libsignal reference implementation (v0.105.4) at the machine-code level. Only ~1% of the codebase needed human review. The team validated the approach by reintroducing two historical bugs and confirming the system correctly flagged them as insecure — a notable proof point for AI-assisted formal verification of security-critical code.

Read at Google DeepMind →
Jensen Huang resigns as Nvidia CEO after affair scandal Nvidia's co-founder and longtime CEO steps down immediately; CFO Colette Kress becomes interim CEO as shares drop 7%. Industry & Trends Nvidia · 2026-11-12

Jensen Huang resigned as President and CEO of Nvidia effective immediately on Nov 12, after admitting to an affair with an alleged Chinese spy and reports of a second affair with a subordinate. CFO Colette Kress was named interim CEO and Harvey Jones became the company's first board chair; a permanent successor search is underway covering internal and external candidates. A major leadership disruption at the company sitting at the center of the AI compute buildout.

Read at Nvidia →
AWS patches SSRF flaw in MCP Gateway Registry CVE-2026-125926 let an authorized user trick the gateway into sending requests to internal network services via a crafted upstream address. AI Security & Safety Cyber Press · 2026-11-11

AWS disclosed and patched CVE-2026-125926, an SSRF vulnerability in mcp-gateway-registry (versions 1.0.3–1.34.0) affecting the server registration/update API. A user with permission to register or modify MCP servers could set a crafted upstream address to make the gateway issue requests to internal network services. Fixed in version 1.35.0. Adds to a growing list of MCP-infrastructure CVEs (ContextForge MCP Gateway, aws-serverless-mcp-server) disclosed this month as MCP gateways proliferate in production.

Read at Cyber Press →
Google launches Gemini 4 Hydrogen Google's new Flash-class model claims 2.5x faster task completion and 5x fewer fabricated claims than its predecessor. Model & Product Releases Google · 2026-11-12

Google released Gemini 4 Hydrogen on Nov 12 as the successor to Gemini 3.8 Flash, rolling out across the Gemini API, AI Studio, Gemini for Business, the Gemini app, and Search. Google claims it completes agentic tasks 2.5x faster and hallucinates more than 5x less often than its predecessor; pricing is $1.50/$7.50 per million input/output tokens. Adds an Antigravity Orchestrator option that can switch models mid-task for research workflows.

Read at Google →
Claude Fable 5.2 and Claude Mythos 5.2 released Anthropic's updated flagship pair targets coding, finance, and scientific work, up to 30% faster than the prior generation. Model & Product Releases Anthropic · 2026-11-09

Anthropic introduced Claude Fable 5.2 (public, safeguarded) and Claude Mythos 5.2 (restricted access, fewer safeguards for vetted users) on Nov 9, pitched as its most capable models yet for technical work in coding, finance, and the sciences, running up to 30% faster than the 5.1 generation. Pricing stays at $8/$40 per million input/output tokens.

Read at Anthropic →
NVIDIA contributes OpenShell agent sandboxing runtime to CNCF A kernel-enforced sandbox for AI agents — Landlock, seccomp, SELinux, namespace isolation — moves to open governance under the CNCF. Tools & Frameworks CNCF · 2026-11-10

NVIDIA, with Red Hat, Docker, and Canonical, contributed OpenShell to the Cloud Native Computing Foundation as a Sandbox project. OpenShell runs each AI agent in a sandbox restricting file/host/credential access via five layers of kernel-enforced controls (Landlock, seccomp, SELinux MAC, user/network namespace isolation) plus per-binary OPA/Rego policy and TLS-intercepting L7 inspection; it's adopting Cedar (another CNCF project) as its policy language. Driven partly by the rise of prompt-injection incidents, it's aimed squarely at practitioners running agents with real tool/filesystem access.

Read at CNCF →
Cloudflare open-sources Kitesurf, its lightweight browser engine for AI agents A Chromium alternative built from Blitz, Stylo, and Boa-in-WASM gives agents web access without a full browser's CPU/memory cost. Tools & Frameworks Cloudflare · 2026-11-12

Cloudflare open-sourced Kitesurf under Apache 2.0 — an experimental browser engine purpose-built for AI agents that stitches together Blitz (HTML/DOM/layout), Stylo (CSS), and Boa (JS-in-WASM fallback) instead of running full Chromium. It exposes a Chrome DevTools Protocol endpoint so existing Puppeteer/agent-harness tooling can connect, and is designed to scale cheaply across Cloudflare Workers. Developers can now self-host it rather than relying only on Cloudflare's managed Browser Rendering API — a practical tool for anyone building web-browsing agents at scale.

Read at Cloudflare →
Anthropic sets IPO price range, pricing confirmed for Nov 18 Shares priced at $70-$78, targeting roughly a $1 trillion valuation; order book already ~6x oversubscribed. Industry & Trends TradingKey · 2026-11-14

Anthropic's amended S-1 set a price range of $70.00-$78.00 for 825 million Class A shares, targeting pricing on Nov 18 and Nasdaq trading under ticker ANT from Nov 19. The deal values the company around $1 trillion ($1.2T fully diluted) — roughly half the $2T figure once floated — with the order book already about 6x oversubscribed. Strategic investors including Microsoft, Nvidia, Alphabet, and Micron are buying roughly $26B of stock at the IPO price. Q3 2026 revenue was reported at ~$17.0B, up from $1.4B a year earlier, against a $567M operating loss.

Read at TradingKey →
OpenAI launches ChatGPT Wallet for agent-initiated purchases ChatGPT's "dots" agents can now pay with tokenized cards via Amex/Mastercard/Visa — never seeing the real card number. Industry & Trends CNBC · 2026-11-13

OpenAI began a limited rollout of ChatGPT Wallet on Nov 13, letting its autonomous "dots" agents complete purchases using a tokenized credential rather than the full card number, built with American Express, Mastercard, and Visa. Initially limited to US ChatGPT Pro subscribers 18+ with eligible US-issued cards; every purchase requires a biometric/security check. OpenAI plans future pre-authorized delegation with spending limits and categories. Notably, Visa testing earlier this year found AI shopping assistants failed to complete purchases correctly roughly 43% of the time — a reliability caveat on top of the new attack surface agent-held payment credentials create.

Read at CNBC →
United Airlines blocks third-party AI agents from booking flights Citing credential-sharing and unintended-purchase risks, United becomes the first major US carrier to formally restrict personal AI agents. Industry & Trends TechCrunch · 2026-11-12

United Airlines updated its booking, ticketing, and distribution policies to restrict personal/third-party AI agents from completing purchases or managing reservations on its platforms, directly and via travel agencies/aggregators. Agents can still look up flights and fares but can't finalize a purchase, and United can cancel without notice any reservation an AI agent created or altered. United cites risks from shared credentials, unintended purchases, MileagePlus misuse, and aviation rules requiring human check-in acknowledgment — the first explicit stand by a US airline against agentic-AI access to its systems.

Read at TechCrunch →
Red Hat and Nvidia's OpenShell driven partly by rising prompt-injection incidents Industry framing piece ties the OpenShell sandbox push directly to the surge of prompt-injection CVEs this year. AI Security & Safety SiliconANGLE · 2026-11-11

Coverage of the OpenShell CNCF contribution notes it's a direct industry response to escalating prompt-injection incidents against agentic systems; separately, trackers now count 52 CVEs in LLM/agent software since January 2026 where prompt injection is the root cause or consequence — a steep rise that underscores why runtime sandboxing for agents is becoming a priority rather than a nice-to-have.

Read at SiliconANGLE →
JailBound: FOL-guided jailbreak evaluation framework accepted at NeurIPS 2026 New benchmark uses first-order-logic-guided dual-branch search to map behavioral boundaries across 46 LLMs from 13 model families. AI Security & Safety arXiv · 2026-11-05

A NeurIPS 2026-accepted paper introduces JailBound, a jailbreak evaluation framework combining automated benchmark construction with first-order-logic-guided dual-branch search, tested across 46 LLMs spanning 13 model families under one unified protocol — a useful reference point for red teams standardizing jailbreak measurement methodology.

Read at arXiv →
LangChain 1.5.0 ships, drops Python 3.10 support The agent-framework update pairs with langchain-core 1.8 as LangChain 0.3/LangGraph 0.4 head toward end-of-maintenance in December. Tools & Frameworks Releasebot / LangChain · 2026-11-11

LangChain released version 1.5.0 on Nov 11, dropping Python 3.10 support alongside langchain-core 1.8. Google's ADK (Python) also shipped 2.14.1 on Nov 12. Both land as LangChain 0.3 and LangGraph 0.4 approach the end of their maintenance window (security patches/critical fixes) in December 2026 — practitioners on older branches should plan migration now.

Read at Releasebot / LangChain →