Daily Brief ↗ source

AI/ML Security & Trends

The dominant story is a one-two punch of AI infrastructure breaches: Fireworks AI disclosed a widened October breach (billing data, secrets, and incomplete model weights exposed via a hacked cloud provider and "LLMjacking") just as NEAR AI disclosed a separate, unrelated RCE breach via a flaw in the open-source SGLang inference engine — both within 48 hours of each other, underscoring how inference-serving infrastructure has become a top attack surface. Anthropic's release of Claude Fable 5.2 and the restricted Claude Mythos 5.2 is the big model news of the day.

16 stories 4 high priority 5 categories
Fireworks AI widens breach disclosure: model weights, secrets, billing data exposed A hacked cloud provider let attackers reach Fireworks R&D machines, stealing up to 7.69% of some customers' model weights. Breaches & Incidents Fireworks AI · 2026-11-09

On Nov 9, Fireworks AI (backed by NVIDIA) disclosed that its October 24 incident was broader than first reported: attackers who breached a third-party cloud provider used that foothold to reach Fireworks development machines, exposing customer account data, billing records, API secrets, datasets, and incomplete model weights (up to 7.69% of weights for some customers) via IAM role chaining. The company ties the activity to financially motivated 'LLMjacking' and has engaged Palo Alto Networks Unit 42 for forensics. This is Fireworks' second disclosed incident in under three weeks, directly impacting AI model confidentiality — a serious model-theft precedent for inference-as-a-service providers.

Read at Fireworks AI →
NEAR AI breached via SGLang RCE vulnerability, 100 users and 18 agents exposed An unauthenticated RCE in the SGLang inference engine let an attacker pivot into NEAR AI's internal systems and legacy agents. Breaches & Incidents The Block · 2026-11-10

On Nov 8, an attacker exploited a remote code execution vulnerability in the open-source SGLang inference engine running on NEAR AI's servers, gaining code execution on one inference server and pivoting to an internal system behind a deprecated Private Chat feature. Roughly 100 users' account details (emails, names, plan/balance info) were downloaded, and 18 legacy AI agents not running inside a Trusted Execution Environment had chat logs, settings, and logs exfiltrated; two customers' API keys were also exposed and revoked. NEAR AI patched SGLang, rotated all credentials, and restored its Cloud service by Nov 10 — a notable case showing TEE/enclave protections don't cover vulnerabilities in software running inside them.

Read at The Block →
Anthropic launches Claude Fable 5.2 and restricted Claude Mythos 5.2 Fable 5.2 gets cheaper and faster; Mythos 5.2 is gated to Life Sciences/Cyber-verified orgs for high-stakes reasoning work. Model & Product Releases Anthropic · 2026-11-09

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Nov 9, ahead of its reported IPO plans. Fable 5.2 drops pricing to $8/$40 per million input/output tokens (from $10/$50), runs up to 30% faster, and improves coding, reasoning, and vision; Mythos 5.2 is a restricted-access model for demanding long-running agentic and reasoning work, limited to organizations in Anthropic's Life Sciences and Cyber verification programs — relevant for security researchers needing vetted access to frontier reasoning capability.

Read at Anthropic →
OpenAI closes $11B funding tranche at $1.25T valuation, below target OpenAI's first tranche fell short of its $1.4T ask amid an AI-stock selloff; more funding expected in coming weeks. Industry & Trends Tech Funding News · 2026-11-06

OpenAI closed the first portion of a new funding round on Nov 6, receiving $11 billion at a $1.25 trillion valuation — below the $1.4 trillion the company had originally sought as part of a planned $30B+ raise. The round may grow to $20 billion total in coming weeks. The lowered ask comes amid a broader selloff in public AI stocks, a signal worth watching for capital availability across the sector.

Read at Tech Funding News →
Mendix Studio Pro MCP server exposes HTTP listener beyond loopback (CVSS 8.8) Mendix's bundled MCP Server fails to restrict its listener to localhost, letting network attackers interact with it remotely. AI Security & Safety Strix Security · 2026-11-10

CVE-2026-112136, published Nov 10, affects Mendix Studio Pro 11's bundled MCP Server, which does not properly restrict its HTTP listener to the loopback interface, potentially exposing it to remote systems. Attackers with network access could interact with the MCP server directly, bypassing the intended local-only trust boundary — another entry in the fast-growing catalogue of MCP implementations shipping with insecure default network exposure.

Read at Strix Security →
AtlasAiDev MCP Server for WordPress has privilege-escalation flaw (CVSS 8.8) Incorrect privilege assignment in a WordPress MCP server allows escalation, published just a day before the Mendix MCP flaw. AI Security & Safety Strix Security · 2026-11-09

CVE-2026-121576, published Nov 9, describes an incorrect privilege assignment vulnerability in the AtlasAiDev MCP Server for WordPress, rated CVSS 8.8, allowing privilege escalation. Combined with the Mendix MCP flaw disclosed the next day and NEAR AI's SGLang RCE, the week shows a cluster of serious vulnerabilities across the AI-agent/MCP supply chain landing almost simultaneously.

Read at Strix Security →
Tracker: 52 prompt-injection CVEs filed against LLM/agent software in 2026 NVD data shows prompt injection is rarely the whole bug — it's usually the entry point to classic flaws like SSRF. AI Security & Safety DEV Community · 2026-11-06

An analysis of the National Vulnerability Database found 52 CVEs since January 2026 in LLM and agent software where prompt injection is a direct cause or consequence, as of Nov 6. The most common associated weakness is CWE-918 (server-side request forgery, 7 cases) followed by CWE-436 (interpretation conflict, 5 cases), reinforcing that prompt injection typically functions as an entry point into traditional vulnerability classes rather than standing alone — useful data for anyone building threat models for agentic systems.

Read at DEV Community →
DeepSeek V4.1 Flash 'Uncensored' model proliferates on red-team/TEE platforms A weights-level decensored DeepSeek variant hits 100% answered on HarmBench, now served inside TEEs for red-teaming use. AI Security & Safety LocallyUncensored · 2026-11-05

DeepSeek V4.1 Flash 1017 Uncensored — a community (dealignai) fine-tune with refusals removed at the weight level — went live on NEAR AI Cloud and other providers in early November, explicitly marketed for red-teaming and security research and served inside a Trusted Execution Environment so findings stay private. The variant reportedly answers 100% of HarmBench prompts versus 1.6% for the stock model, illustrating how easily safety alignment can be stripped from open-weight models and redistributed — a direct dual-use concern for AI-safety researchers.

Read at LocallyUncensored →
Meta releases Muse Ember, a $0.10/M-token agent model with 1M context Meta's new small model targets high-volume agentic sub-tasks, claiming Haiku 5.5-level intelligence at roughly a third the cost. Model & Product Releases MarkTechPost · 2026-11-09

Meta launched Muse Ember 1.0 on Nov 9/10, a small multimodal model (text, image, audio, video) aimed at tool invocation, coding sub-tasks, routing, and long-document parsing for agent workloads. It offers a 1M-token context window, 512K max output, and pricing starting at $0.10/M input and $0.50/M output tokens. Meta claims it scores comparably to Claude Haiku 5.5 on the Artificial Analysis Intelligence Index at roughly a third of the cost per task — a direct play for high-volume agentic pipelines.

Read at MarkTechPost →
OpenAI brings ChatGPT Enterprise to AWS Marketplace Enterprises can now buy ChatGPT Enterprise through AWS, drawing down existing cloud spend commitments. Industry & Trends OpenAI · 2026-11-09

Starting Nov 9, enterprises can purchase ChatGPT Enterprise directly through AWS Marketplace, using existing cloud contracts and unified billing while drawing down pre-committed AWS cloud spend — opening a new commercial distribution channel for OpenAI and tying its enterprise revenue more closely to AWS's go-to-market infrastructure, notable given OpenAI's existing Microsoft relationship.

Read at OpenAI →
Shield AI to acquire counter-drone radar maker Echodyne for up to $5.3B Shield AI's largest-ever acquisition pairs autonomous systems with counter-drone radar at a $28.8B valuation. Industry & Trends DroneXL · 2026-11-09

Shield AI agreed on Nov 5 to acquire Echodyne, a counter-drone radar maker, for up to $5.3 billion while raising $2.5 billion at a $28.8 billion valuation — Shield AI's fifth and largest acquisition to date, consolidating autonomous defense-AI and sensing hardware under one roof.

Read at DroneXL →
@tsps/mcp-server denial-of-service flaw lets unauthenticated remote memory exhaustion A high-severity DoS CVE lets attackers crash Node.js MCP servers via unbounded memory growth on the /mcp endpoint. AI Security & Safety DailyCVE · 2026-11-09

CVE-2026-63655, disclosed in early November, describes a denial-of-service flaw in the @tsps/mcp-server package's HTTP server mode: an unauthenticated remote attacker can cause the Node.js server process to consume unbounded memory via POST requests to the /mcp endpoint, crashing the service. It's part of a broader pattern of MCP implementations shipping without basic resource-exhaustion protections.

Read at DailyCVE →
Google Workspace rolls out custom 'Skills' for Gemini Reusable custom instructions let teams bake their own rules, templates, and reference files into Gemini's behavior. Tools & Frameworks Google Workspace Updates · 2026-11-09

Google began rolling out 'Skills' in Google Workspace on Nov 9 — reusable, custom instruction sets that guide Gemini to apply a team's specific rules, templates, and reference files across tasks, available first to Rapid Release domains. This is Google's answer to the growing practitioner trend (also seen in Claude and OpenAI tooling) of packaging reusable agent behaviors rather than re-prompting from scratch each time.

Read at Google Workspace Updates →
xAI adds default invisible watermarking to Grok API models xAI begins statistically watermarking Grok API outputs by default, citing EU AI Act transparency obligations. Industry & Trends Crypto Briefing · 2026-11-07

Starting Nov 7, xAI added invisible statistical watermarking across its Grok API models by default, aligning with EU AI Act transparency requirements and following similar moves already made by Anthropic. The change affects provenance/detectability of AI-generated text at scale and is relevant to anyone building abuse-detection or content-authenticity tooling.

Read at Crypto Briefing →
Google begins sunsetting Gemini Code Assist entitlements, Pro model access tightens Google is unwinding bundled Gemini Code Assist entitlements and restricting Pro-model access to paid tiers starting Nov 13. Industry & Trends Google Cloud · 2026-11-09

Google began rolling out the end of bundled Gemini Code Assist entitlements for existing Gemini Enterprise Standard/Plus subscribers outside the US/EU/global regions on Nov 9, with full removal by March 31, 2027. Separately, from Nov 13, Google AI Plus subscribers lose the ability to select Gemini Pro models in the Gemini app and AI Mode in Search, restricting Pro access to Google AI Pro subscribers and above — a monetization tightening move across Google's AI product line.

Read at Google Cloud →
Future Secure AI raises $80M, becomes Australia's newest unicorn A three-year-old Australian AI-security startup hits a $1.2B valuation via convertible notes. Industry & Trends EntrepreneurLoop · 2026-11-06

Future Secure AI raised $80 million on Nov 6 through convertible notes at a $1.2 billion valuation, making the three-year-old firm Australia's newest tech unicorn — notable as one of the larger recent raises specifically in the AI-security vendor space outside the US.

Read at EntrepreneurLoop →