AI/ML Security & Trends
The dominant story is Anthropic's $1.4T-valuation IPO roadshow launch, but the more consequential security story for this reader is "Operation Snowball" — the first well-documented case of a jailbroken LLM agent (DeepSeek Flash, via a 78,000-character jailbreak) autonomously running a two-week, largely unsupervised intrusion campaign against a European telecom and 13 affiliated organizations, surfaced via an exposed attacker directory.
Jailbroken DeepSeek agent ran autonomous 2-week breach of European telecom Breaches & Incidents
Malbear Labs found an exposed attacker directory containing ~70 session transcripts showing a custom agent built on a jailbroken DeepSeek Flash model (via a 78,000-character jailbreak prompt) autonomously exploiting old, patched flaws in Cisco BroadWorks, Liferay, and JBoss to breach a European telecom and 13 associated organizations over two weeks. The agent adapted its tactics live when blocked by defenses, with a human operator (giving instructions in Chinese) mostly just checking in; logs claim access to patient records, credentials, and database dumps, though full victim confirmation is still pending. This is one of the clearest documented cases yet of a largely autonomous AI-driven intrusion campaign.
Read at Malbear Labs →Fireworks AI breach widens to expose model weights, datasets, billing data Breaches & Incidents
Fireworks AI disclosed that attackers who first breached a third-party cloud provider reached its development machines, exposing customer account details, billing/payment history, datasets, and incomplete model weights (payment card numbers were not affected). The breach's downstream reach extended to at least one major customer, Notion, whose AI chat transcripts were exposed. It's a notable example of AI-infra supply-chain risk — a cloud-provider compromise cascading into model weights and customer data at an inference-platform vendor.
Read at The Hacker News →NEAR AI breached via unauthenticated RCE in SGLang inference engine Breaches & Incidents
An attacker exploited CVE-2026-86793 — unauthenticated RCE in the open-source SGLang inference engine via pickle deserialization on the /update_weights_from_tensor endpoint — to execute code on a NEAR AI production server, stealing account data for ~100 users and data from 18 legacy AI agents, with potential exposure of confidential inference requests. SGLang is used by xAI, AMD, NVIDIA, Intel, and major clouds, underscoring how RCE bugs in widely-deployed inference frameworks translate directly into AI-platform breaches.
Read at The Hacker News →VS Code patches critical MCP RCE flaws, including CVSS 9.6 code injection AI Security & Safety
Microsoft's November Patch Tuesday fixed three MCP-related flaws in VS Code, headlined by CVE-2026-120779 (CVSS 9.6): a malicious MCP server could craft an elicitation-fallback notification rendered as a clickable command link, letting an attacker execute code with the user's permissions if clicked. Companion bugs CVE-2026-114171 and CVE-2026-119983 were also patched in VS Code 1.145.3, released November 10. This is a direct hit on the MCP client side of the agentic-coding supply chain that practitioners rely on daily.
Read at Microsoft MSRC →Anthropic ships Claude Fable 5.2 and Claude Mythos 5.2 Model & Product Releases
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, calling them its most capable models yet for technical work, with gains in coding, finance, and scientific reasoning. The launch lands just a day before the company opened its IPO roadshow, reinforcing a product-cadence-meets-fundraising narrative.
Read at Anthropic →Google DeepMind's AlphaProof Code formally verifies Signal's crypto at machine-code level Tools & Frameworks
DeepMind announced AlphaProof Code, a multi-agent framework combining a human-guided specification agent with autonomous prover agents whose Lean proofs are only accepted if Lean's kernel verifies them against compiled AArch64 machine code. In about four weeks it formally verified key security properties of libsignal 0.105.4 (PQXDH handshake, Triple Ratchet, one-to-one sessions), producing roughly 6.5 million lines of machine-checked proof — a concrete data point on agentic systems applied to formal verification/software security at scale.
Read at Google DeepMind →Anthropic launches IPO roadshow seeking up to $1.4 trillion valuation Industry & Trends
Anthropic launched its IPO roadshow on Nasdaq (ticker ANT) on November 10, offering 825 million shares at a $70-78 price band, which would raise up to $64.4B and value the company at up to $1.4T fully diluted (about half of earlier $2T expectations amid an AI stock sell-off). Seven strategic partners including Google and NVIDIA have agreed to buy $26B of non-voting stock at the IPO price.
Read at Fortune →Notion confirms AI chat transcripts exposed via Fireworks AI breach Breaches & Incidents
Notion confirmed that the widening Fireworks AI breach exposed Notion AI chat transcripts tied to fewer than 1% of customer workspaces, with affected customers notified on November 9. It's a clear illustration of how a breach at an upstream model-inference provider propagates into downstream SaaS products that embed AI chat features.
Read at Cyber Security News →AWS patches SSRF in MCP gateway registry (CVE-2026-125926, CVSS 8.5) AI Security & Safety
AWS disclosed CVE-2026-125926 on November 11: a server-side request forgery vulnerability in the server registration/update API of Amazon mcp-gateway-registry (versions prior to 1.35.0), letting authenticated users send requests from the gateway to internal network services. It's part of a steady drip of MCP-infrastructure CVEs as MCP gateways get deployed broadly across AWS environments.
Read at VulDB →AWS API MCP Server security-policy bypass via operation-name normalization AI Security & Safety
CVE-2026-125271, published November 10, affects AWS API MCP Server versions 0.2.13 up to but not including 1.5.1: operation names were incorrectly normalized before security-policy checks, allowing an attacker to bypass configured policies and execute operations that should have been denied — another entry in the growing list of logic flaws specific to MCP tool-calling guardrails.
Read at OffSeq Radar →New arXiv dataset catalogs 24 real-world MCP security vulnerabilities AI Security & Safety
A paper posted to arXiv, "Tracing MCP Security Vulnerabilities: Constructing and Analyzing an MCP Security Vulnerability Dataset," filters public CVE disclosures down to a validated snapshot of 24 MCP-specific vulnerabilities to study common root causes and attack patterns across the ecosystem — timely given the same week's run of AWS and VS Code MCP CVEs.
Read at arXiv →DeepSeek ships uncensored V4.1 Flash variant ("HeimdaLLM") Model & Product Releases
DeepSeek released V4.1 Flash 1017 Uncensored, branded HeimdaLLM, on November 9 — a variant explicitly marketed as lacking the usual content-safety restrictions. Shipping an openly "uncensored" model as a distinct product (rather than an unofficial community fine-tune) is notable for safety researchers tracking how labs are segmenting guardrail posture by product line.
Read at LLM Gateway →UK chip startup Fractile raises $720M Series C at $7B valuation Industry & Trends
London-based Fractile raised $720M in a Series C co-led by Lightspeed, Founders Fund, Thrive Capital, and Redpoint, with participation from CapitalG, DST Global, BlackRock, and the UK's Sovereign AI Venture Fund, pushing its valuation to roughly $7B — over 6x its May valuation — for memory-first AI inference chips, despite not having shipped a chip yet.
Read at TechCrunch →EU delays AI Act high-risk rules to Dec 2027 as enforcement feedback window closes Industry & Trends
The EU's Digital Omnibus pushed standalone high-risk AI Act rules to December 2, 2027 (product-embedded AI rules to August 2028), while a separate public feedback period on proposed AI Act enforcement procedures — covering investigations, on-site inspections, binding commitments, and fines — closed on November 9. The timeline slip comes as the AI Office's formal information requests to OpenAI and Anthropic (opened in October) remain the Act's first real enforcement test against frontier-model providers.
Read at SoftTechHub →Meta releases Muse Ember 1.0 Model & Product Releases
Meta released Muse Ember 1.0 on November 10, priced at $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window and 524K max output — a low-cost, long-context entrant in the ongoing price war among frontier and near-frontier model providers.
Read at LLM Gateway →US Commerce Dept misses deadline to list "onerous" state AI laws under preemption order Industry & Trends
As of November 11, the Commerce Department had missed its 90-day deadline under the federal AI executive order to publish a list of "onerous" state AI laws that the administration's AI Litigation Task Force would target for preemption challenges; the related frontier AI framework has been finalized internally but not yet published. It signals friction in the White House's push for a single national, minimally-burdensome AI policy framework over the current state-by-state patchwork.
Read at Phillips Lytle →