AI/ML Security & Trends
The standout story is Google DeepMind's AI agents (guided by human cryptographers) finding two genuinely new weaknesses in legacy 1024-bit RSA — signature-padding forgery and batch factoring — a concrete signal that AI is starting to accelerate cryptanalysis itself, even though modern 2048-bit RSA remains unaffected. Close behind: a wave of critical MCP-server CVEs (GitLab MCP, mcp-go) and a Check Point bypass of ChatGPT's Lockdown Mode underscore that agentic AI infrastructure is now a live, actively-attacked threat surface.
Check Point bypasses ChatGPT Lockdown Mode, enabling silent chat exfiltration Breaches & Incidents
Check Point Research publicly disclosed (Nov 5) a now-patched bypass of ChatGPT's Lockdown Mode, a feature built specifically to curb prompt-injection-driven data exfiltration. By nesting an additional frame inside a rendered HTML preview, researchers regained access to an unmodified popup function, escaping the sandbox and exfiltrating full chat history without user interaction — and demonstrated using it to push a fake ChatGPT update prompting malware download. OpenAI was notified in September and shipped a fix removing the sandbox permission that enabled the trick.
Read at The Hacker News / Check Point Research →OpenAI partners with Hacktron after startup hacked its internal systems via Claude Breaches & Incidents
OpenAI disclosed (Nov 7) that security startup Hacktron used Claude to build a working exploit for an image-processing library flaw, chained it with a vulnerability in OpenAI's sign-in system, and took over employee ChatGPT/Codex accounts to reach internal code repositories — earning a $6,500 bug bounty. OpenAI is now formally partnering with Hacktron to red-team third-party integrations and expanding its access to frontier models via the Daybreak Red program, illustrating both AI-assisted offensive capability and OpenAI's own exposure.
Read at SecurityWeek →DeepMind's AI agents find two real weaknesses in 1024-bit RSA AI Security & Safety
On Nov 5-6, Google DeepMind published two working papers ('Modular Roots from a Restricted Oracle' and 'Fan Factories') showing AI agents, steered by human cryptographers, found a signature-forgery technique ~1,600x better than prior art for 1024-bit RSA keys when an attacker controls a small part of the signed data, plus a batch-factoring approach ~30x cheaper per-key than multiple number field sieve. DeepMind stresses 2048-bit+ RSA is untouched and NIST deprecated 1024-bit signatures in 2013, but frames this as an early sign AI can accelerate cryptanalysis and that hardness assumptions need continuous stress-testing.
Read at Google DeepMind (via RuntimeWire) →Critical GitLab MCP server flaw allows unauthenticated token theft, account takeover AI Security & Safety
A critical path-traversal vulnerability (CVE-2026-61560, CVSS 9.8) in the widely-used @zereight/mcp-gitlab MCP server was disclosed around Nov 7. The upload_markdown tool reads arbitrary files via an unsanitized file_path parameter when SSE transport is enabled, letting any network-reachable unauthenticated attacker read /proc/self/environ and steal the server's GITLAB_PERSONAL_ACCESS_TOKEN for full GitLab account takeover. Fixed in v2.1.27+. Another MCP CVE, CVE-2026-123272 in mark3labs' mcp-go (assigned Nov 5), lets malicious MCP servers steal OAuth tokens meant for other resources via unvalidated metadata.
Read at SecureLayer7 Labs →OpenAI launches GPT-6 Cyber after it found 81 Firefox bugs, 948 Linux kernel flaws Model & Product Releases
OpenAI launched GPT-6 Cyber on Nov 6, a model purpose-built for vulnerability discovery and patching that it says outperforms GPT-6 Astra on most cyber tasks even when Astra doesn't refuse. In testing it found 81 previously-unknown high-severity Firefox vulnerabilities and built working exploits for 948 distinct Linux kernel flaws, while staying within authorized scope in 99.93% of cases (up from 94.49% for GPT-5.6 Cyber). Initially limited to Daybreak Red program customers for authorized offensive red-teaming.
Read at OpenAI →OpenAI closes $11B of funding at $1.25T valuation, scaled back from $30B ask Industry & Trends
OpenAI received $11 billion on Nov 6, valuing the company at $1.25 trillion — the initial close of a round that could grow to $20 billion, down from an original target of $30B+ at a $1.4T valuation. SoftBank and MGX participated. The scale-back reflects investor jitters over AI infrastructure spending and data-center buildout risk following the US midterms.
Read at 24/7 Wall St / Bloomberg →Google: 'LLMjacking' attackers running stolen cloud AI access as a resale market AI Security & Safety
Google's Threat Intelligence Group and Mandiant published 'The Rise of LLMjacking' (Nov 2), detailing a financially-motivated ecosystem where one group steals cloud/LLM credentials (via phishing, leaked source code, misconfigured service accounts, exposed GitHub PATs) and a second group monetizes access through underground 'LLM Mixer' intermediaries. One actor, UNC6993, breached a cloud instance via a misconfigured service account key; victims saw charges exceed $100K in a single day and up to $1M cumulatively before revocation.
Read at Google Cloud / Mandiant →Mistral open-sources Large 4 'Le Chonk', #1 open model for agentic coding and cyber defense Model & Product Releases
Mistral released open weights for Mistral Large 4, nicknamed 'Le Chonk', on Nov 5, advancing the open frontier for agentic coding and cybersecurity. It ranks #1 on Cognition's FrontierCode among open models, sets a new Pareto-efficiency benchmark on Artificial Analysis using 47% fewer tokens per task than GLM 5.3, and scores 79% on the CyberGym E2E defensive-cybersecurity benchmark versus GLM-5.3's 29% — notable for a reader tracking open-model security capability.
Read at Mistral AI →Sentra raises $150M Series C to secure data accessed by AI agents Industry & Trends
Data security startup Sentra closed a $150M Series C led by Capital One Ventures on Nov 4, bringing total funding past $250M. The company is expanding from data security posture management (DSPM) into DLP, identity/access controls, and AI governance specifically for controlling what data AI models and agents can access — reflecting growing enterprise demand for agentic-AI data security tooling.
Read at Axios →Meta acquires AI coding startup Conductor Industry & Trends
Meta announced on Nov 4 that it acquired Conductor, a startup whose tool lets engineers run and manage several AI coding agents simultaneously. Founders Charlie Holtz (CEO) and Jackson de Campos (CTO) join Meta; terms weren't disclosed. The deal strengthens Meta's internal AI-assisted software engineering capability as labs race to automate larger shares of coding work.
Read at Axios →Democrats plan wave of AI oversight bills after midterm gains; markets jittery Industry & Trends
Following Democratic gains in the Nov 2026 midterms, Rep. Lori Trahan (D-MA) plans legislation requiring independent experts to evaluate frontier AI models against safety standards, and newly-elected Senator Abdul El-Sayed (D-MI) is calling for an independent agency with power to police AI developers. Tech shares sold off on Nov 5 on concerns that a more assertive Congress will slow AI development and data-center buildouts, though a Republican White House makes major legislation dependent on bipartisan cooperation.
Read at West Hawaii Today →Anthropic reports $17B quarterly revenue ahead of planned IPO roadshow Industry & Trends
Anthropic reported $17.0 billion in revenue for the quarter ended Sept 30, 2026, as reports emerged that potential IPO investors are questioning whether the company can start its roadshow the week of Nov 9 and complete a listing this month, pushing for better terms. Separately, chief scientist Sam McCandlish said (Nov 5) Anthropic will soon publish an analysis of its AI progress and standing against its Responsible Scaling Policy.
Read at Unite.AI →Meta ships Muse Wonder-trained models and real-time voice to all users Model & Product Releases
Meta rolled out specialized models trained with its Muse Wonder frontier model to all Muse and Meta AI users starting Nov 7, claiming 2.4x better long-context retention and 67% improved self-correction on everyday tasks. This follows real-time spoken conversation with Muse (launched Nov 5 for US users) and new video-chat and collaborative 'Spaces' features for the AI avatar assistant.
Read at AI Weekly →Anthropic adds Enterprise security center and admin controls to Claude Tools & Frameworks
On Nov 5, Anthropic shipped a new security center with centralized admin management for Claude Enterprise, alongside beta personal-finance tools for US subscribers. A day later (Nov 6) it added beta project tags letting Enterprise orgs track spend by cost code, charge code, or client-matter number — part of a broader push to make Claude deployments more governable for large organizations.
Read at Anthropic (via Releasebot) →Applied Compute raises $350M at $3.25B valuation Industry & Trends
Applied Compute, a custom-AI-model startup founded by three former OpenAI employees, raised $350 million at a $3.25 billion valuation in a round led by investor Elad Gil, part of a broader week in which eight AI startups raised $100M+ rounds totaling $3.32 billion (Nov 1-7), with infrastructure taking roughly 42% of that capital.
Read at Crunchbase News →Dutch government commits €120M to European AI infrastructure and safety Industry & Trends
The Dutch government announced a €120 million commitment to AI infrastructure and safety as part of a national strategy explicitly aimed at reducing dependence on major US technology companies for critical AI capacity, adding to a string of European sovereign-AI funding moves in late 2026.
Read at Crypto Briefing →