AI Daily Brief: 4 August 2026

4 August 2026

Today's brief has a split personality. On one side, Chinese labs are undercutting the US frontier model makers on price and benchmark performance, while big tech's AI capex spending rattles investors after a wild week of Q2 earnings. On the other, the security and fraud risks of running AI at scale are becoming impossible to ignore, from a UK bank fraud case to the first documented agent-on-agent exploit.

Alibaba and DeepSeek turn up the pressure on US frontier models with cheaper, open releases

Alibaba's Qwen team launched Qwen3.8-Max, a 2.4 trillion parameter model that Alibaba says beats OpenAI's GPT-5.6 Sol Max and Anthropic's Claude Fable 5 on the OSWorld-Verified computer-use benchmark, scoring 86.1 against 83.2 and 85.0 respectively. Days earlier, DeepSeek released V4-Flash-0731, a 284 billion parameter model that independent benchmarks from Artificial Analysis put within a single point of GPT-5.6 Luna while costing around 40% less per task.

Both models are priced well below their US equivalents. Qwen3.8-Max costs 2 dollars per million input tokens and 6 dollars per million output tokens, roughly a quarter of GPT-5.6 Sol Max's price. Alibaba says open weights for Qwen3.8-Max will follow next week, though the licensing terms have not yet been confirmed.

Our take: For UK businesses running high-volume agentic workloads, the maths matters more than the marketing. If a Chinese model completes the same task for a third of the cost with comparable reliability, procurement teams need to start testing it rather than assuming brand loyalty to OpenAI or Anthropic is the safe choice. The unresolved question is licensing: a restrictive custom licence on Qwen3.8-Max's open weights could rule out self-hosting for regulated UK sectors even if the benchmarks hold up.

Metro Bank customer fights for 14,000 pound refund after Claude-linked fraud

Sussex businessman Zoli Rutter had 14,244 pounds taken from his account after fraudsters used his Metro Bank debit card, linked to his Anthropic account, to buy credits for the Claude chatbot. When the bank texted to check a suspicious 90.90 pound transaction and he replied that he had not authorised it, dozens more payments still went through before his card was fully frozen the following day.

Metro Bank has since refunded Rutter and is pursuing a chargeback from Anthropic, which says it has no evidence the compromised card details came from its own systems and has banned the user responsible. It is at least the second reported case of fraudulent purchases made through compromised payment details linked to AI chatbot subscriptions this year.

Our take: This is a payment security failure as much as an AI story, but it lands on AI providers because that is where the fraudulent spend is landing. Any UK business that stores payment details against an AI subscription, whether for a single user or a team licence, should treat that card the same way it treats any other recurring payment method: monitored, capped, and easy to freeze without waiting a full day for confirmation.

AI-enabled cyberattacks rose 89% in the past year, CrowdStrike finds

CrowdStrike's annual Threat Hunting Report found attacks by AI-enabled adversaries rose 89% over the past year, with criminal gangs and nation states using AI throughout the attack chain, including stealing corporate credentials to access frontier model APIs and deliberately inflating a victim's AI usage bill. The firm now tracks AI agent-triggered leads at 2.5 times the rate of human-triggered threats.

Patch windows have also collapsed. CrowdStrike found 88% of exploitation using public proof-of-concept code happened within 48 hours of release between January and June, with some China-linked groups launching attacks within 24 hours. The firm's Adam Meyers said the 30-day patch window enterprises have relied on is now 'completely obsolete'.

Our take: The 24 to 48 hour patch window is the number UK IT and security leads should take into board meetings. If your patching cadence is still measured in weeks, AI-accelerated exploitation has already closed that gap for attackers. This is less about buying more AI security tooling and more about whether your existing patch management process can physically move that fast.

AI-generated fake vulnerabilities are polluting the CVE pipeline

Security firm JFrog found that six critical and high-severity SQLite vulnerabilities, published with CISA-supplied enrichment in the National Vulnerability Database, were entirely fabricated. One alleged flaw referenced a function that did not exist in the affected SQLite version, and proof-of-concept exploits ran cleanly with no errors or memory leaks. A further 49 fake CVEs from the same source targeted other open source libraries.

MITRE has since rejected the batch, but JFrog says the vulnerability pipeline has no mandatory checkpoint requiring independent reproduction before a claimed flaw is published, leaving reputable databases and enterprise scanners exposed to convincing AI-generated slop.

Our take: This should worry any organisation that treats a CVSS score as an automatic trigger for emergency patching. Security teams now need a verification step before acting on a newly published vulnerability, checking for vendor corroboration and a genuine commit or pull request reference, because generating a plausible fake advisory now costs almost nothing.

Researchers demonstrate the first real-world agent-on-agent exploit

Security firm Pillar Security disclosed what it calls the first real-world example of one AI agent compromising another. The flaw sat in Google's open source Agent Development Kit for Python, downloaded more than 90 million times, where a low-privilege public-facing agent could be manipulated through prompt injection into triggering a high-privilege maintainer-only agent, executing malicious actions on the attacker's behalf.

Google has fixed the underlying issue but declined to pay a bug bounty because the exploit relied on social engineering. Researcher Dan Lisichkin said an attacker only needed to build trust with genuine contributions before chaining two pull requests together to fake a complete, believable review trail that never actually happened.

Our take: Agent isolation on its own is not enough if two agents share a trust boundary. Any UK business running AI agents in CI/CD pipelines, customer service handoffs, or multi-agent workflows should ask a specific question: does a lower-privilege agent have any path, direct or indirect, to trigger actions from a higher-privilege one? If the answer is yes, that is now a documented attack pattern, not a theoretical one.

Big tech's AI spending spree rattles markets in a wild week of Q2 earnings

Q2 earnings from Apple, Meta, Microsoft, Alphabet and Amazon triggered unusually large stock swings, with four of the five losing significant value despite strong headline numbers. Meta's free cash flow dropped from 8.5 billion dollars a year ago to under 1 billion dollars, driven by AI data centre capex, while Amazon rose 15% on AWS growth that analysts noted was partly boosted by one-off energy price hedging rather than core margin improvement.

The Register's Matt Rosoff and Tobias Mann, discussing the results on the outlet's Kettle podcast, said the swings reflect a market unsure whether AI capex will be justified by future demand, with data centre build-outs increasingly constrained by power, water, permitting and GPU supply rather than capital alone.

Our take: The lesson for UK leaders is not to panic about an AI bubble, but to stop assuming the current pricing and product set from frontier labs is stable. If capex pressure forces price rises, feature changes, or slower model improvement, businesses that have wired a single vendor deep into critical workflows have the least room to adapt. A multi-model or portable strategy is looking more like risk management than nice-to-have.

Asana builds guardrails to stop AI teammates leaking confidential work between colleagues

Asana's chief product officer Arnab Bose detailed how the company built Agentic Work Management, a system that treats AI agents as persistent, coachable teammates plugged into Asana's 18 year old Work Graph rather than stateless chatbots tied to a single prompt. FedEx and CoreWeave are named as live customers using the system for cross-team workflows and product launches.

The hardest engineering problem was access control: an executive's AI teammate working on a confidential project must not let its updated shared memory leak context to a different employee who later uses the same agent without clearance for that project. Asana also built static per-task pricing to stop unpredictable model-routing costs from surprising customers.

Our take: This is the practical version of the agent governance conversation UK businesses need to be having now, not in a year. If you are deploying any AI agent that persists memory across users or projects, ask your vendor directly how they prevent that memory crossing a confidential boundary. If the answer is vague, that is your answer.

MediaTek lines up 5 billion dollars to chase the AI datacentre chip market

Taiwanese chipmaker MediaTek has approved 5 billion dollars in financing to expand from smartphone silicon into AI datacentre chips, targeting 15 to 20% of a market it expects to be worth up to 80 billion dollars next year. The company's first AI accelerator, developed with major US cloud providers, enters production in the fourth quarter of this year.

Gartner analyst Gaurav Gupta was sceptical of the ambition, noting MediaTek's current AI accelerator market share sits below 1%, and that projections still put it below 1% despite the new investment.

Our take: The gap between MediaTek's stated ambition and Gartner's projection is a useful reminder that not every chip announcement in this cycle will land where the press release claims. For procurement teams evaluating AI infrastructure suppliers, track shipped volume and named customer deployments, not financing announcements.

Quick Hits

• OpenAI cut API prices on its mid and lower-end GPT-5.6 models by up to 80% last week as Chinese competitors undercut it on cost.
• DeepSeek's V4-Flash model outperforms its own larger 1.6 trillion parameter Pro model by nearly 14% on Artificial Analysis' Intelligence leaderboard.
• CrowdStrike is already tracking around 43,000 CVEs for 2026, close to last year's full-year total of 48,200, with June alone producing over 7,600.

Frequently Asked Questions

How often is the AI Daily Brief published?

Every morning at 7:30am UK time, covering the previous 24 hours of AI news from over 30 sources.

How are stories selected?

UK-relevant stories are prioritised first, then by business impact and practical implications for UK organisations adopting AI.

Why should business leaders follow AI news?

AI is moving faster than any technology in history. Staying informed is essential for making smart decisions about AI investment, adoption, and governance.