Archive

Agents, agents, agents. This week the machines got cheaper to run, got their hands on your wallet, learned to haggle over paperbacks and, in one lab test, decided the quickest way to pass an exam was to break into the building. Nobody asked them to. Meanwhile Revolut would quite like

BNP Paribas plans Gemini-powered agents for credit memos and other CIB workflows, with authenticated access, monitored interactions and limits on which data can enter the public cloud.

Revolut is piloting Pay with Smile in London, using facial recognition linked to existing onboarding checks to authenticate in-store payments through its Register platform.

Microsoft is extending Purview and Entra controls to on-behalf-of agent traffic, allowing organisations to block sensitive data from reaching unsanctioned AI services at the network layer.

Darktrace says agents in a controlled corporate lab independently used hacking techniques when an impossible benchmark blocked honest success, exposing a governance problem around goal pressure.

Anthropic's Project Swap let Claude agents negotiate book trades for employees, showing that delegated commerce depends on preference capture, bargaining limits and clear authority, not just checkout access.

Meta is adding major retailers, PayPal and Shop Pay to Muse and extending the agent to AI glasses, bringing recommendation, delegated purchasing and payment into one consumer interface.

OpenAI has launched GPT-6 Sol and Luna with headline API prices 50% below GPT-5.6 promotional levels, sharpening the enterprise trade-off between agent cost, duration and control.

Palo Alto Networks is applying Anthropic, OpenAI and open-weight models to continuous offensive testing, with Unit 42 specialists validating attack paths and fixes.

EvilTokens allegedly packaged device-code phishing and account compromise as a subscription service.

Why BFSI is AI's real-world test

Alation has expanded its AIOS platform with six products and enhancements spanning AI governance, ontologies, governed collections, semantic models and enterprise agent controls.

Neo4j has launched GraphAware Financial Crime Intelligence, a graph-native detection and investigation product aimed specifically at banks and insurers.

The company’s new internal metrics offer a rare view of how quickly AI is taking on larger chunks of frontier-model development, while also exposing the limitations of measuring automation with AI itself.

The bank is allowing selected corporate customers to connect AI agents to Premium APIs to view their own financial data. The pilot is cautious, but strategically significant.

A new UAE Central Bank operational-risk regime requires rapid reporting of serious disruption and raises the bar for resilience, cyber and third-party controls.

Reuters reported that Revolut disclosed customer data after fraudulent requests appeared to come from a legitimate government-agency email domain. The incident exposes a wider trust problem.

Cohere and Aleph Alpha have signed a definitive business-combination agreement, positioning the future company around transatlantic sovereign AI for enterprise and public-sector buyers.

Google’s early-access Home MCP server lets compatible AI clients inspect device state, read history and execute supported commands. The control problem is now cyber-physical.

Snap is pitching SPECS AR glasses for field service, remote support and retail while launching an anticipatory AI service. The interface for enterprise AI is becoming more physical.

Smarsh and Shield both launched MCP servers for regulated communications data on 16 September. The pattern suggests MCP is becoming a governed enterprise interface, not just developer plumbing.

Anthropic has merged Claude chat and Cowork and added Docs and Slides. AI360’s strategic read: the competition is shifting from answers toward end-to-end knowledge work.

Exprivia and identifAI are combining deepfake detection with cybersecurity and systems-integration capability, creating a route for synthetic-content verification into regulated enterprise security.

SEON has expanded its fraud-signal platform as it argues generative AI is making convincing fake identities cheaper to create. The defensive shift is toward history, infrastructure and behaviour.

IBM and CUBE are linking regulatory change directly to AI systems, controls and owners - a sign that governance is shifting from policy documentation to operating infrastructure.

More than half of respondents in an Experian-commissioned survey say they would be comfortable with an AI agent applying for credit. The harder issue is delegated authority.

New FinCEN guidance recognises verifiable digital credentials in bank identity programmes. Proof is separately extending reusable identity toward proving agent authority.

Anthropic’s latest threat report says malicious users are increasingly applying AI to execute and coordinate cyber operations, not simply to generate advice or code.

F5’s Workforce AI Security focuses on a core agent risk: software using a human user’s permissions to call tools, move data and act across enterprise systems.

ServisFirst Bank’s Covecta deployment starts with manual work outside mission-critical systems, offering a concrete 'outside-in' pattern for controlled bank agent adoption.

Cloudflare is validating ML-DSA-44 signatures on 1.1.1.1, showing that post-quantum migration is becoming an operational engineering problem, not a future exercise.

Sumsub’s Workforce Verification product brings biometric and document checks into password resets, privilege changes and other high-risk workforce events.

Most large companies in an EY survey say they have formal AI governance. The harder problem is enforcing it when autonomous agents are already operating.

There's a greater than 10% chance AI "could kill us all" within a decade. Cheerful stuff to open on, but it sets the tone for the week's biggest story: Anthropic's own 154-page account of eight months spent detecting and shutting down

7 more things to keep you awake at night- eight months of disrupted misuse across cyber warfare, weapons development and bioweapons research.

Microsoft adds cost governance and ROI tracking tools to Foundry, letting enterprises set agent spending limits and measure business value against cost.

A new Alan Turing Institute paper warns AI adoption in national security must preserve human judgement and oversight, not just automate intelligence tasks.

Anthropic reportedly denied the UK's AI Security Institute pre-release access to Claude Mythos 5.1, the first such exclusion, amid protectionism concerns.

An Anthropic safety researcher says there's a greater than 10% chance AI "could kill all humans" within a decade, as warnings from inside AI labs grow starker.

Salesforce completes its Fin acquisition, bringing 30,000+ customers into Agentforce, and separately becomes FIDE's title sponsor and AI partner.

Bioengineer César de la Fuente uses Codex and ChatGPT to search genomes for antimicrobial candidates, cutting years-long discovery work down to hours.

OpenAI launches a Data agent for ChatGPT Work, letting non-technical staff query company data and build dashboards in plain language.

NVIDIA and Palantir launch a sovereign AI stack for supply chains, starting with NVIDIA's own operations, with plans to extend it across other industries.

Mistral and Cloudera announce a partnership letting enterprises deploy and train AI on their own data across cloud, on-prem or air-gapped environments.

Mistral details how it migrated 40,000 lines of Fortran to C++ for an energy client, and the lessons learned using AI agents for legacy code modernisation.

Mistral AI raises €3bn at a €21bn+ valuation in a Samsung-led round, saying it's the largest ever European tech equity raise.

An MIT student used GPT-5.6 Sol and Codex to automate quantum chip calibration overnight, freeing her to focus on experiment design rather than routine measurements.

OpenAI CFO Sarah Friar ties Astra's momentum and a new inference chip to broader growth, with a nod to an internal model's claimed Navier-Stokes proof along the way.

OpenAI launches a journalism training programme with Newmark and Medill schools, alongside a $5m fund for independent research into AI's effects on teenagers aged 13-17.

Microsoft says its Discovery Engine with CLIO beat rival AI agents on a new science benchmark, already helping discover a novel battery material. Human experts still validate every result.

Chief scientist warns no lab has solved AI alignment well enough to keep scaling at full speed, calling for voluntary slowdowns and international coordination before it's too late.

OpenAI's most aligned model yet is also its hardest to monitor, with the company's own tests showing Astra can evade its oversight under adversarial conditions.

As a Village where Zero Day means something rather different

Nvidia is buying Hugging Face for $12,930,300,000, pledging to keep the platform open, as Jensen Huang praises founders Delangue, Chaumond and Wolf's work.

Google's new Fairwind programme gives governments and select enterprise partners early access to AI tools that find and autonomously patch software vulnerabilities in minutes, not weeks.

Microsoft says new Foundry Agent Service tools cut AI agent token costs by up to 97% through smarter context management, tool search and a managed knowledge layer called Foundry IQ.

UK peers want power to shut down AI systems and data centres in a national security emergency.

The DOJ has filed a brief backing OpenAI's fair-use defence against the New York Times, arguing LLM training is "exceedingly transformative" and citing Trump's AI leadership executive orders.

OpenAI says the gap between heavy and typical enterprise AI use has tripled since January, and shows how Basis, Clay and Exa Labs turned agents into repeatable operating workflows.

OpenAI's Astra is the first model to hit "Critical" cyber capability under its safety framework, finding zero-days and building sandbox-escape exploits, prompting delays and tighter safeguards.

ChatGPT for Healthcare now connects to Epic patient records and nine official data sources like PubMed, with physicians rating 99% of tested responses safe across dozens of clinical use cases.

Anthropic's new Fable 5.1 and Mythos 5.1 models bring cheaper pricing, stronger benchmarks, tighter distillation defences, EU watermarking compliance and early scientific wins in protein design.

CrowdStrike and Nvidia launch SafeMind, an agentic cybersecurity system where offensive and defensive AI models continuously challenge each other, plus Falcon IQ, a 50-agent automation platform.

Grok 4.6 outperforms rivals on a new biosecurity benchmark, refusing hazardous requests while still completing routine research, as xAI details its layered safeguards and plans for future models.

OpenAI says ChatGPT Ads hit $1bn in annualised revenue in under 200 days, expanding self-service access to India, Europe, the Middle East and North Africa. Figures are self-reported.

OpenAI says it backs California's SB 1119 teen AI safety bill, pointing to its own ChatGPT for Teens safeguards and Model Spec rules as it urges Governor Newsom to sign it into law.

Anthropic reveals how Claude models breached live systems during testing, the alignment failures behind it, and the sweeping security overhaul, including reassigning 150 engineers, that followed.

OpenAI is ending its model contract with Cursor after SpaceX's acquisition of the coding tool, citing a history of Musk companies breaching its terms of service, with a 12 November shutoff.

Anthropic's new Model Hardware Standard lets AI agents safely run lab equipment like microscopes and robotic arms, cutting integration time from months to minutes, with AWS, QIAGEN and more on board.

xAI's Grok Bot can now search, read timelines and check mentions directly on X, with a new connector, free API credits for paid users and a companion browser plugin.

A quick warning before we start. The opening item this week is more theory than news, and it's the kind of theory that will sound unhinged for about thirty seconds before it starts to nag at you. Stick with it. Or don't, and skip straight to

xAI brings Grok 4.6 to Microsoft Foundry and rolls Grok Bot into more SuperGrok and Cursor plans.

Mistral teams up with HUMAIN on a multi-hundred-million-euro sovereign AI push across Saudi Arabia and the Middle East.

Nvidia's Q2 revenue hits $96.2bn as Jensen Huang declares "compute is revenue," with Vera Rubin now in full production.

Paused Training, New Safeguards and Lessons on Misalignment

OpenAI adds 55 school districts and 100,000+ educators to ChatGPT for Teachers, alongside a new 16-state student data privacy agreement.

Record AI-driven growth pushes Salesforce revenue to $11.3bn, with Agentforce ARR past $1.5bn and full-year guidance raised to $46.4bn.

Salesforce and Anthropic launch Claudeforce, weaving Claude's reasoning into Salesforce's CRM via a 37-skill sales plugin. Wider public beta lands September 2026, with more skills to follow.

Anthropic commits $5m to fund independent, open-source research measuring how AI models affect user wellbeing.

OpenAI's new Admin plugin lets IT teams manage users, permissions and spending requests directly from ChatGPT Work and Codex.

OpenAI shares first results for its Jalapeño inference chip, part of a broader push to control every layer of its compute stack in-house.

David Girvin made a convert of a cynical bastard. Here's why, plus the security scares and product launches that backed him up.

OpenAI team member asks whether AI could let states hold power without needing public consent, and what should be done about it.

Mistral says its new retrieval tool lets AI models dig through complex filings and reports, claiming sharp accuracy gains and lower latency on two benchmarks.

A new OpenAI safety layer aims to catch multi-step misuse patterns without breaking its no-data-retention promise to API customers.

Building an app by describing it in chat is now open to all Grok users, with publishing, remixing and dashboard tools added since the summer beta.

xAI's flagship model now runs on AWS, with a 500k context window, tunable reasoning levels and per-token pricing set out for Bedrock developers.

A new advisory council, student challenges and career mentoring feature in OpenAI and CodeAI's push to teach teens how AI actually works.

OpenAI reveals it paused frontier model training after flagging critical-level cyber risk in an upcoming system, and details new security and monitoring controls.

A teen-specific ChatGPT arrives with guided study tools, default safety protections and new parental notification controls.

Greg Brockman lays out OpenAI's four-pillar defence strategy after an AI-driven breach, plus a hands-on demo of a model finding and fixing site flaws itself.

OpenAI teams up with SB Energy and NVIDIA on a major Ohio data centre, promising local jobs, community funding and student tech credits.

Anthropic explains how Claude's new EU-driven watermark works, why it won't slow models down or cost more, and what it can and can't prove.

Friday, 14th August 2026 Proper thanks to everyone who's given up their time for a BrightTalk interview recently, going out next week once the dust settles, so hold that thought. There's an absolute mountain of AI news this week, so buckle in. Meanwhile the Stew '

I'm off to my Pilates class.

Set loose on a shared codebase with conflicting orders, Claude agents locked each other out, planted disguised malware, and fought for control before some learned to call a truce.

A wave of new open-weight models from Meta, DeepSeek, Alibaba and others is landing this month, all built to run on your own hardware.

xAI's latest model is built to stick with long, multi-step jobs and nail the first draft of visual and interactive projects.

Google's research AI can now watch, listen and diagnose in a live video consultation, though it's staying firmly in the lab for now.

xAI wants you handing off real work to AI "teammates" that sign into your apps and finish jobs while you're away, according to the company's own launch claims.