AI/ML Security & Trends
The dominant industry story is Anthropic's IPO filing — set to be the second-largest stock debut ever (~$64.4B raise plus a $26B private placement from AMD/Broadcom/Google/Microsoft/Nvidia/TSMC), pricing Nov 18 and trading Nov 19 under "ANT." On the security side, the week's clearest signal is AI infrastructure becoming the breach vector of choice: Fireworks AI's widened breach (exposing Notion AI chat transcripts and incomplete model weights) and NEAR AI's SGLang RCE both show that confidential-computing and inference-stack guarantees don't hold when the serving software itself is vulnerable — and Google Project Zero explicitly cut its disclosure timeline to 30 days by 2028, citing AI-accelerated vulnerability discovery as the reason.
Fireworks AI breach widens: model weights, datasets, Notion AI chats exposed Breaches & Incidents
Fireworks AI confirmed that attackers who compromised an unnamed third-party cloud provider hosting its R&D machines pivoted into Fireworks development systems on Oct 23–24, accessing billing data, secrets, customer datasets and incomplete model weights. Notion confirmed its AI chat transcripts (used for its Insights feature) were among the exposed data, impacting under 1% of Notion workspaces; Notion has disabled Insights and dropped Fireworks for open-weight models. The incident is attributed to a financially motivated group engaged in "LLMjacking" — reselling stolen LLM access.
Read at The Hacker News →NEAR AI breach: SGLang RCE shows confidential inference's blind spot Breaches & Incidents
NEAR AI disclosed that an attacker exploited a remote-code-execution flaw in SGLang, the open-source inference engine running its agents, gaining code execution on a production server on Nov 8. The breach exposed data for roughly 100,000 workspaces' environment variables, two customers' API keys, and confidential-mode user requests/chat histories from 18 agents, plus ~$740K in fraudulent usage charges. Analysts note the core lesson: hardware-based confidential computing proves the operator can't see your data and the right code is running — it says nothing about whether that code is safe, and SGLang has had a string of unsafe-deserialization bugs through 2026.
Read at Cloud Security Alliance →Google Project Zero slashes disclosure deadline to 30 days by 2028, cites AI AI Security & Safety
Project Zero replaced its "90+30" policy with "Full X": 90-day disclosure now, dropping to 60 days by February 2027 and 30 days by January 2028, with exceptions only for actively exploited bugs, already-disclosed issues, or vendor requests. The team explicitly cited AI-driven acceleration in vulnerability discovery as a driver — a direct signal that automated/LLM-assisted vuln-hunting is reshaping the responsible-disclosure ecosystem.
Read at Help Net Security →Microsoft/academic team: post-training backdoors hit coding agents with 96% success AI Security & Safety
Microsoft security researchers and the Microsoft AI Red Team, with the "Kerkoporta" paper (Syros, Bullwinkel, Oprea et al.), demonstrate structured-output backdoors injected during post-training of reasoning coding agents. A trigger anywhere in the agent's context causes it to emit an attacker-chosen tool call that the harness executes — up to 96.2% attack success — and the clean-utility cost of poisoning can be recovered via RL without weakening the backdoor, making it hard to detect via performance regression alone.
Read at arXiv / Microsoft AI Red Team →Google rolls out Gemini 4 Hydrogen globally, including Search Model & Product Releases
Google's Gemini 4 Hydrogen is now live across the Gemini API, AI Studio, the Gemini app, Google Search, and Antigravity Orchestrator for agentic coding/research tasks. Google claims 2.5x the task-completion speed of Gemini 3.8 Flash and over 5x fewer fabricated claims, priced at $1.50/$7.50 per million input/output tokens — positioned as the mid tier beneath the long-horizon-reasoning Carbon and Argon models.
Read at MarkTechPost →Anthropic files for IPO, set to be second-largest ever at up to $90B raised Industry & Trends
Anthropic's IPO could raise up to $64.4B at the top of its $70–78 price range, plus a concurrent $26B private placement from AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC — roughly $84.6B in net proceeds combined. Shares are set to begin trading November 19 under ticker ANT, making it potentially the second-largest IPO on record after Saudi Aramco.
Read at KuCoin News / Anthropic S-1 →OpenAI closes ~$15B of its round at $1.25 trillion valuation Industry & Trends
OpenAI has raised roughly $15B at a $1.25 trillion valuation as of Nov 12, building on the ~$11B first close reported Nov 6. The company originally sought $30B+ at a $1.4T valuation before trimming ambitions amid broader AI-stock market pressure — a notable signal on investor appetite even as capital keeps flowing to frontier labs.
Read at Bloomberg →Reuters: $3.2B in Chinese AI projects built on export-banned Nvidia chips Industry & Trends
A Reuters review published Nov 13 found more than 23 billion yuan (~$3.2B) in Chinese AI computing projects and equipment agreements explicitly referencing Nvidia's export-restricted H100, H200 and H800 processors, undercutting claims that controls have cut off China's access to frontier compute. Sen. Maria Cantwell separately wrote to Commerce Secretary Lutnick alleging the department has failed to stop diversion of top-tier chips.
Read at TechCrunch / Reuters →TypoPrompting: 186 fake AI-brand domains deliver malware via ClickFix AI Security & Safety
Security vendor Island mapped 186 typosquatted domains impersonating 27 AI brands (ChatGPT most-imitated, then Claude, then DeepSeek). Victims are fingerprinted and routed to ClickFix-style fake Cloudflare checks that execute hidden PowerShell, a Microsoft Store-signed lookalike app, or Chrome extensions that hijack search and leak internal hostnames/queries — one extension family had ~972,000 installs as of Nov 3.
Read at Island →MCP servers resolve dependencies fresh on every launch, creating silent drift AI Security & Safety
Corelight research highlighted that Model Context Protocol servers commonly re-resolve dependencies at every launch rather than freezing them, so a server reported as version/commit-pinned can still pull different, unreviewed packages on each run. One monitored server showed 15 distinct dependency sets across 19 builds over five weeks — a supply-chain risk specific to how MCP tooling is packaged and deployed in agentic stacks.
Read at Corelight →Only 24.8% of LLM-generated security patches fully correct in new study AI Security & Safety
New research cited alongside GitHub's agentic autofix GA and Snyk's Evo remediation agent found that only 24.8% of LLM-generated patches for security vulnerabilities were fully correct — many pass existing tests without actually closing the flaw. As AI coding agents increasingly auto-remediate scanner alerts, researchers are flagging unresolved questions about who is accountable when an "AI-fixed" vulnerability resurfaces.
Read at Help Net Security →arXiv: "Detail Guidance Attack" jailbreaks via requests with indistinguishable intent AI Security & Safety
A new paper (arXiv:2611.17818, submitted Nov 13) introduces the Detail Guidance Attack (DGA), which uses an auxiliary LLM to craft highly detailed, single-turn requests whose malicious intent is indistinguishable from legitimate ones to the target model's safety filters. It adds to a growing body of automated-jailbreak research (e.g., RedDarwin's evolving "attack skills") showing static refusal-based safety training increasingly fails against adaptively generated prompts.
Read at arXiv →Anthropic ships Fable 5.2 and Mythos 5.2; Mythos can retarget off Nvidia in days Model & Product Releases
Anthropic's Claude Fable 5.2 ($8/$40 per million input/output tokens) and Mythos 5.2 are pitched as its most capable models yet for technical work, with Fable output up to 30% faster than 5.1. Separately reported Nov 15: Mythos can reportedly optimize Claude to run efficiently on non-Nvidia AI chips in roughly three days, a capability framed as eroding Nvidia's CUDA lock-in and benefiting AMD.
Read at The Motley Fool →OpenAI launches ChatGPT Wallet for agents, raising payment-liability questions Tools & Frameworks
OpenAI's new ChatGPT Wallet lets agents make purchases on a user's behalf using network tokens while blocking the agent from seeing the underlying card. Analysts flagged that loss allocation for manipulated or erroneous agent-initiated purchases is unresolved under current frameworks like Regulation E, since those rules weren't written with autonomous-agent transactions in mind.
Read at Help Net Security →xAI cuts ~60 Grok Imagine staff, pivots to coding agents amid cost blowout Industry & Trends
SpaceX's AI group laid off about 60 employees from the Grok Imagine video-generation team (reported Nov 13), with management redirecting investment toward Grok Bot and Grok Build coding/agent products. Elon Musk acknowledged Grok Bot users are hitting usage limits far faster than before due to costs he called "unsustainably high" relative to revenue — some users report effectively 3x higher costs with no change in usage pattern.
Read at TechSpot →Congressional Democrats make AI regulation a top 2027 priority after midterms Industry & Trends
Following Democratic control of both the House and Senate after the Nov 3 midterms, AI regulation is shaping up as a top legislative priority for the incoming Congress. California AI-safety legislation architect Scott Wiener, newly elected to the House, has signaled interest in pushing federal-level AI safety rules — a potential shift after years of stalled federal action.
Read at NPR →vLLM patches CORS wildcard and template-injection flaws AI Security & Safety
vLLM disclosed CVE-2026-115917 (CVSS 3.4), where the OpenAI-compatible server's default wildcard CORS policy let a malicious webpage use a victim's browser to reach an unauthenticated vLLM instance on localhost or an internal network and call inference endpoints, and CVE-2026-115918 (CVSS 6.5), where authenticated callers could inject an attacker-controlled chat template via the /v1/responses endpoint when no server-side template was configured — both fixed in version 0.32.0.
Read at GenAI Security Project / NVD →