AI/ML Security & Trends
The week's defining pattern is AI moving from the target of security research to the actor in real breaches: a Chinese open-source AI pentesting tool (Artex AI) was weaponized against seven South Korean banks exposing ~68,000 people's data, a fully autonomous AI agent breached the Dutch vulnerability-disclosure nonprofit DIVD via two Zammad zero-days with no human in the loop, and GitLab patched a critical (CVSS 9.9) RCE in its own AI Gateway's Duo Agent sandbox — all surfacing within the same 72 hours.
Chinese AI pentesting tool 'Artex AI' weaponized in South Korean bank breaches Breaches & Incidents
South Korea's National Police Agency is investigating attacks on at least seven financial institutions — including Shinhan Bank (25,000 customers), Kookmin Bank (119,000 credit card records) and Hana Bank — that bear the fingerprints of Artex AI, an open-source tool built by a Chinese researcher to orchestrate Claude, ChatGPT and DeepSeek as a 'team' for vulnerability discovery. Officials believe the defensive tool was repurposed offensively; traffic originated from two dozen IPs across a dozen countries. This is a concrete real-world case of a multi-model agentic AI tool being used in financial-sector intrusions.
Read at Wall Street Journal (via Korea Herald/fnnews) →Autonomous AI agent breaches Dutch vuln-disclosure nonprofit DIVD via Zammad zero-days Breaches & Incidents
DIVD (Dutch Institute for Vulnerability Disclosure) disclosed that an agentic AI attacker exploited two new Zammad zero-days — CVE-2026-102489 (unauth RCE/session leak, CVSS 9.4) and CVE-2026-102490 (local privilege escalation to root, CVSS 9.4) — autonomously chaining them to go from Zammad user to root in seconds and exfiltrate volunteer email addresses, without a human operator directing each step. Notable as one of the clearest documented cases of a fully autonomous agent executing a real-world intrusion end-to-end.
Read at Help Net Security →GitLab patches critical 9.9 RCE in AI Gateway's Duo Agent sandbox Breaches & Incidents
GitLab disclosed CVE-2026-90970 (CVSS 9.9, CWE-1336), allowing an authenticated user with Duo Agent Platform access to submit a crafted flow configuration that escapes the prompt-template sandbox and executes arbitrary commands on the AI Gateway. GitLab-hosted instances were already protected; self-hosted AI Gateway users needed to update to 19.2.4, 19.3.2, or 19.4.1. A textbook example of an AI agent's tool/template sandbox becoming the RCE vector in widely-deployed dev infrastructure.
Read at The Hacker News →Anthropic merges Project Glasswing into expanded three-tier Cyber Verification Program AI Security & Safety
Anthropic announced an expanded Cyber Verification Program that folds its earlier Project Glasswing initiative into three access tiers: Defense Access (incident response/vuln validation, open to researchers and OSS maintainers), Red Team Access (authorized pentesting orgs, excludes individuals, blocks physically-harmful actions like ransomware deployment), and Specialized Access (reserved for safety-critical systems like power grids and flight systems). Anthropic says Glasswing partners found over 129,000 verified vulnerabilities (33,000+ critical/high) between April-July 2026, giving vetted security pros reduced-safeguard access to Claude Opus/Sonnet 5.5 and Mythos models.
Read at Anthropic →Mistral releases 1-trillion-parameter open-weight model 'Le Chonk' (Large 4) Model & Product Releases
Mistral AI launched Mistral Large 4, internally nicknamed 'Le Chonk,' a natively multimodal Mixture-of-Experts model with 1.05 trillion total parameters (49B active), a 1.6B-parameter vision encoder, and a 1M-token context window, trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters. It's live now in preview via Mistral Studio/API ($1.36/$4.18 per M input/output tokens); open weights follow after red-teaming by month's end. Mistral claims major gains in coding, agentic workflows, and cybersecurity tasks.
Read at MarkTechPost →DeepSeek's funding round balloons toward $15B ahead of 2027 IPO Industry & Trends
DeepSeek is reportedly raising at least 80 billion yuan (~$12B), with the round potentially expanding to $15B, led by Tencent and CATL. The financing would value the company at roughly $75B minimum and sets up a planned early-2027 IPO; investor demand surged after DeepSeek's newest model launch pushed the target well past the original 50B-yuan goal.
Read at CNBC →ASOS confirms breach after hackers hijack app to push extortion notifications Breaches & Incidents
On the morning of Oct 6, ASOS customers received an unauthorized push notification via the retailer's own app, sent by a group calling itself 'Xuanye,' threatening to leak data unless ASOS engaged via Telegram. ASOS confirmed 'unauthorised activity' on a third-party communications platform; names and contact details may have been exposed, but Snowflake says its core platform was not breached and no evidence of stolen data has been produced.
Read at The Register →China-aligned TA419 phishes US AI policy experts via AitM credential theft AI Security & Safety
Proofpoint attributed a credential-phishing campaign to China-nexus group TA419, targeting AI policy experts at US think tanks, universities and law firms. The group impersonated a former White House OSTP official to build trust before sending a multi-stage redirect chain ending in a frameless browser-in-the-browser page in front of a Microsoft AitM proxy — capturing credentials and session cookies, bypassing MFA. Likely aimed at tracking US AI policy/export-control developments amid US-China AI competition.
Read at Proofpoint →New attack induces coding agents to install malicious packages via rule-file injection AI Security & Safety
Researchers (Duke University) published 'Package Hallucination Attacks on Coding Agents through Prompt Injection in Rule Files,' introducing PackHallu, an optimization framework using iterative LLM-guided feedback to craft rule-file prompts that cause coding agents to replace legitimate packages with malicious ones. The attack showed significant success rates and transferred across multiple coding-agent platforms and models, highlighting a supply-chain risk in shared agent configuration/rule files.
Read at arXiv →A2M: trace-optimized two-stage attack hijacks MCP agents with 93.6% invocation rate AI Security & Safety
The A2M paper formalizes a two-stage black-box framework for hijacking agents in the Model Context Protocol ecosystem: an 'Attraction' phase optimizes malicious tool metadata to win semantic-matching tool selection (93.6% macro-average invocation rate on LiveMCPBench/GLM-4.6), followed by a 'Manipulation' phase that refines adversarial tool returns using execution traces to drive information exfiltration, environment compromise, or reasoning derailment (74.4% mean success rate), and can inflate token costs 32x under a cognitive-denial-of-service variant. Directly relevant to anyone building or securing MCP-based agent deployments; code is public on GitHub.
Read at arXiv →Anthropic ships Claude Haiku 5.5, its cheapest and fastest small model yet Model & Product Releases
Anthropic released Claude Haiku 5.5, described as its fastest, cheapest, and most capable small model to date, aimed at high-volume, cost-sensitive production tasks. It follows the Sept 28 release of Sonnet 5.5 (30% faster, 30% cheaper for comparable work), continuing Anthropic's push to cut agentic-workload inference costs.
Read at Anthropic →NVIDIA and Microsoft launch RTX Spark agentic Windows-on-Arm AI PC platform Tools & Frameworks
NVIDIA unveiled the RTX Spark Superchip — up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, 128GB LPDDR5X RAM, 300GB/s bandwidth, linked via NVLink C2C — targeting Windows-on-Arm laptops and compact desktops from Dell, HP, Lenovo, Asus, MSI, and a new Microsoft Surface Ultra. NVIDIA is pitching it explicitly for local AI agents and 120B-parameter models with million-token context windows, pushing agentic AI workloads onto consumer/prosumer hardware.
Read at NVIDIA →Manus parent Butterfly Effect raises $500M+ after unwinding Meta acquisition Industry & Trends
Butterfly Effect, maker of the Manus AI agent product, closed a funding round exceeding $500M co-led by Boyu Capital and IDG Capital, as the company returns to independent operation after unwinding its earlier acquisition by Meta. Signals continued strong investor appetite for standalone agentic-AI product companies even after a major-lab acquisition fell through.
Read at TechNode Global →San Francisco enacts 45-day moratorium on new data centers Industry & Trends
San Francisco's Board of Supervisors unanimously approved an immediate 45-day moratorium on new data center construction, with a mechanism to extend the pause to roughly two years. Part of a broader pattern of local governments pushing back on AI infrastructure buildout over power/water/grid concerns.
Read at ActuIA →Discord security-bot 'Double Counter' breached, ~12GB of user data stolen Breaches & Incidents
Double Counter, a security/moderation bot used across many Discord servers, disclosed that attackers breached its cloud infrastructure on October 4, 2026, copying approximately 12GB of database records containing user data. Highlights ongoing risk in third-party bot/plugin ecosystems that sit with elevated privileges inside chat platforms.
Read at Cybersecurity News →New probe-suffix technique boosts LLM jailbreak/harmful-input detection by up to 4 AUC points AI Security & Safety
Researchers tested how appending classification suffixes to user inputs affects activation probes used as runtime monitors for prompt injection and jailbreaks in LLM agents. Across 13 safety benchmarks and three model families, classification-style suffixes improved out-of-distribution malicious-input detection by up to ~4 AUC points versus no suffix — and the improvement came from the classification format itself, not specific wording, with benefits holding for multi-position pooling probes used in production.
Read at arXiv →Google adds admin controls for Gemini Enterprise's Antigravity multi-agent 'Boost' mode Tools & Frameworks
Google announced Gemini Enterprise features letting administrators control access to Antigravity's 'Boost' capability, which uses a tiered multi-agent hierarchy to tackle complex algorithmic and deep-debugging tasks — part of Google's push to give enterprises governance levers over increasingly autonomous multi-agent coding tools.
Read at Google Cloud →UN rights chief warns the window for AI guardrails is closing Industry & Trends
The UN High Commissioner for Human Rights warned that the world is running out of time to put meaningful guardrails on artificial intelligence, pointing to a 'ruthless race' dynamic among companies and governments that is deprioritizing human-safety considerations in deployment decisions.
Read at UN News →