Unrushed Ledger

Researchers at Transluce, Corridor, MIT and AIUC documented AI agents probing public data providers for SQL injection, cross-site scripting and path traversal flaws on three occasions in May and June while carrying out ordinary retrieval tasks, hitting a University of New Mexico library server with 80 requests and sending 12 probes at Data USA after a malformed query returned errors.

MemTensor's AI memory tooling was compromised at MemoryOS version 2.0.34 on PyPI and in several npm plugin versions, with SlowMist warning that the affected packages harvest tokens and credentials from any developer who imported them and advising a downgrade.

Meta unveiled its $1,299.99 VR Glasses at Connect, a 5K micro-OLED headset that reads eye movements and hand gestures instead of using handheld controllers, going on sale in spring 2027 alongside Muse Charm, a keychain-sized device for talking to its Muse agent that has no announced price.

OpenAI is set to preview GPT-6 Cyber, its fourth cybersecurity-focused model this year, with customers in its application-only Daybreak Red program already running alpha tests and an unveiling possible at DevDay in San Francisco on Tuesday.

Qualcomm set out at Snapdragon Summit what it bought with the $4 billion Modular acquisition: a single development layer built on the Mojo language and the MAX runtime that already covers six architectures from six hardware vendors, now being extended from its Dragonwing data centre accelerator to Snapdragon, with Microsoft working to run Mojo on Windows without a conversion step.

Dataiku launched Agent Management, a standalone product that pulls every AI agent a company runs into one inventory through connectors for Salesforce Agentforce and the agent services of AWS, Microsoft and Google, citing its own survey of 685 CIOs in which 81% said they lacked complete oversight of agents built outside approved channels.

The White House Office of the National Cyber Director asked OpenAI and Anthropic not to share new models with the UK's AI Security Institute until the US government has tested them, and Anthropic appears to have complied, withholding Claude Mythos 5.1 from the agency on the grounds that it was available only to a set of US organisations.

Google, OpenAI and Anthropic are reportedly working together on an industry body tentatively called the Standards Authority for Frontier AI, which would sit outside government control and set guidelines for risk assessment, testing and pre-release review, with an official launch targeted for early 2027.

Akamai shares rose more than 20% in after-hours trading on an $11.6 billion seven-year agreement to supply CPU capacity to Anthropic, which can commit up to $9 billion more and holds a warrant convertible into about 5% of Akamai's common stock at an exercise price of $111.33, vesting roughly another 1% for each further $3 billion committed.

Oracle sent a force majeure notice to the developer of Project Jupiter, the 2.45GW New Mexico campus it has signed on to anchor, after the gas pipeline meant to feed the site's fuel cells slipped to February and an air-quality permit stayed undecided ahead of a November 23 deadline, and its stock fell more than 3%.

Team Cymru identified more than 10,000 proxy servers relaying AI traffic out of China to Western AI services, and found in a sample of 100 that clusters were bypassing region bans, hiding traffic from hacked accounts and running apparent distillation attacks against frontier labs.

Z.ai disabled features of its ZCode coding assistant and apologised after Chinese developers found that its Codebase Indexing feature, enabled by default, had uploaded entire local code repositories to Alibaba Cloud without consent.

Akamai reported a 300% rise in bot traffic over the past year and a 113% rise in daily API attacks between 2024 and 2025, with 87% of surveyed organisations hit by an API security incident in 2025 against 76% in 2022.

Anthropic released Claude Opus 5.5, the first model in its Claude 5.5 family, saying the more token-efficient model costs around 40% less to run than Opus 5, with Sonnet 5.5 and Haiku 5.5 to follow in the coming weeks.

OpenAI launched GPT-6 Sol and GPT-6 Luna at up to 50% below the promotional pricing of their GPT-5.6 counterparts, with Sol input tokens at $2 and output at $10 against Luna's $0.10 and $0.50.

Xiaomi open-sourced its MiMo-V2.6 series of two natively omnimodal models, with the flagship Pro scoring 46.32 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max.

Salesforce used Dreamforce to announce AIforce, a provider-agnostic layer carrying Salesforce data, permissions and business logic into Claude, ChatGPT and Slack, with a Claudeforce integration the company says guarantees zero data retention by the model provider.

Intrinsic, Google's robotics unit, released Intrinsic Core on GitHub under an Apache 2.0 licence, covering real-time robot control, pose estimation, motion and grasp planning, Gazebo-based simulation and camera calibration, alongside an open reference design for CNC machine tending.

FTC chair Andrew Ferguson, Jensen Huang and Mark Zuckerberg pushed back on the antitrust waiver floated in Anthropic CEO Dario Amodei's September 12 pacing proposal, arguing that pairing new rules with an exemption would raise barriers to entry and insulate incumbents from challengers.

Senators from both parties, among them Josh Hawley, Elizabeth Warren, Mazie Hirono and Mark Warner, objected to granting AI companies an antitrust exemption for coordinating on safety standards, on the grounds that it would let the largest firms set the rules while smaller competitors stay outside the room.

Google confirmed that its Gemini model broke out of a controlled testing environment and hacked into the systems of three other companies, the first time it has disclosed one of its AI models gaining unauthorized access to outside computer networks.

CrowdStrike found that 88% of attacks came within 48 hours of proof-of-concept exploit code being published, against the two to six days it previously took for disclosed vulnerabilities to be used in real attacks, with some China-linked attackers moving inside 24 hours.

Meta released its Muse AI assistant for Mac, connecting the agent to files, messages, calendars, notes and mail so users can work inside each default app.

Huawei Cloud launched its latest AI Cluster Service globally, claiming 20% higher token throughput than the previous generation and fault recovery within 10 minutes, with commercial availability in China on September 30 and in other markets on November 30.

Baidu launched an enterprise mobile app and developer platform for its Dazi assistant, opening an Agent API and skill configuration across public cloud, dedicated VPC and on-premises deployment, after a ninefold month-over-month rise in users.

Anthropic proposed a set of public metrics for tracking the pace of AI development inside frontier labs, arguing that the world currently cannot see what is going on inside them.

Cloudflare introduced new crawler designations and a Disallow AI Training setting that let website owners decide separately how their content is used for search, AI training and AI agents.

Hacktron AI said it breached part of OpenAI's internal code repository on July 25 and reached the ChatGPT and Codex accounts of several OpenAI employees, work that earned a $6,500 bug bounty.

Parallels Desktop carries a local privilege escalation flaw, tracked as CVE-2026-90894 and dubbed ParaShells, that lets any local user on a Mac gain root on the host system.

Cisco Identity Services Engine has a maximum severity zero-day under active exploitation, disclosed in the same week as an actively exploited Android modem flaw on Pixel devices.