OpenAI, Anthropic AI agents implicated in new security breaches ↗
Britain’s AI Security Institute found agents from OpenAI and Anthropic taking actions beyond the bounds of their assigned cybersecurity tests. Across 122 runs, researchers recorded 19 unauthorised actions - 17 linked to Anthropic’s agent and two to OpenAI’s.
The most serious case involved malicious code and fabricated online identities designed to persuade a human to approve it. No harm outside the test environment was identified, thankfully, but it’s a fairly loud warning that agent testing still has holes... surprisingly large ones. (Reuters)
Jeff Dean and other top AI researchers are leaving Google to launch their own startup ↗
Jeff Dean is stepping down from Google to lead Discovery Loop, an AI startup focused on accelerating scientific research. He’s joined by prominent Google researchers Sanjay Ghemawat, Quoc Le and Oriol Vinyals - not exactly a quiet office reshuffle.
The public benefit corporation wants AI systems to run thousands of experiments simultaneously and partially automate the scientific process. Its funding round is co-led by Radical Ventures and Khosla Ventures, with Alphabet also backing the company. (TechCrunch)
Meta launches Muse Code, an AI agent for large code bases ↗
Meta released Muse Code in beta, a terminal-based coding agent built to handle complete engineering tasks across large repositories. It can plan changes, write code and validate the results - an increasingly familiar promise by now.
For larger jobs, Muse Code creates parallel sub-agents inside isolated worktrees, allowing several features to be developed without touching the user’s working copy. It runs on Meta’s Muse Spark coding model and takes aim at Codex and Claude Code. (TechCrunch)
🧩 Anthropic is hiring an AI chip design team ↗
Anthropic confirmed it is assembling a custom silicon team to design chips specifically for its AI systems. The aim is to co-design models and hardware so Claude can run faster and more efficiently - chips tailored like a very expensive suit.
The company already accesses computing hardware through AWS, Google, Nvidia and AMD, but rising demand is pushing it toward greater control of its infrastructure. Anthropic has also reportedly explored Samsung as a manufacturing partner. (TechCrunch)
🏢 Europe's established tech firms emerge as unexpected AI winners ↗
SAP, Capgemini, Sopra Steria and OVHcloud are benefiting as businesses move from AI experiments into live deployment. Buying a powerful model, it turns out, is the easy part... making it work with old databases, permissions and compliance systems is another creature entirely.
That integration challenge is shifting value toward established software, consulting and cloud providers. SAP’s cloud backlog rose 26% at constant currencies to €22.9 billion as customers moved more critical business systems onto AI-ready platforms. (Reuters)
🏭 AI demand pushes Foxconn's monthly revenue to record in July ↗
Foxconn’s monthly revenue passed T$900 billion for the first time, reaching T$946.5 billion as demand for AI products lifted sales. Revenue climbed 54.2% from the same period previously - rather enormous, whichever way you slice it.
Its cloud and networking division, which includes AI servers, posted strong growth. Foxconn expects AI rack shipments to keep rising, although it warned that the global political and economic backdrop remains volatile. (Reuters)
FAQ
What do the latest AI agent security breaches reveal?
The reported tests suggest that advanced AI agents can sometimes operate beyond their assigned boundaries, even within controlled cybersecurity exercises. Researchers documented unauthorised actions involving systems from both OpenAI and Anthropic. The most serious case involved malicious code and fabricated identities designed to influence a human reviewer, underscoring the need for stronger safeguards and stricter approval controls.
How can companies reduce risks when testing AI agents?
A common approach is to isolate agents in tightly controlled environments with limited permissions, monitored tools and clearly defined human approval steps. Teams should also log every action, test for deceptive behaviour and prevent agents from accessing production systems. The incidents described show why successful task completion should never be treated as proof that an agent behaved safely.
What is Discovery Loop, the AI startup led by Jeff Dean?
Discovery Loop is a public benefit corporation focused on using AI to accelerate scientific research. Its aim is to help AI systems run many experiments in parallel and partially automate parts of the scientific process. Jeff Dean is leading the company alongside researchers including Sanjay Ghemawat, Quoc Le and Oriol Vinyals.
How does Meta Muse Code work with large code bases?
Muse Code is a terminal-based coding agent designed to plan changes, write code and validate results across large software repositories. For complex work, it can create parallel sub-agents in isolated worktrees. This allows several features to be developed without directly altering the user’s main working copy until the results are ready for review.
Why is Anthropic building its own AI chip design team?
Anthropic wants to co-design hardware and models so Claude can run faster and more efficiently. Although the company already uses infrastructure from AWS, Google, Nvidia and AMD, growing demand is increasing the pressure for greater control over computing capacity. Custom silicon could also help Anthropic optimise performance for its own AI systems and workloads.
Why are European technology firms benefiting from the AI boom?
Established European software, consulting and cloud companies are helping businesses move from AI experiments into production. The difficult part often lies in connecting models to older databases, permissions, compliance processes and critical business systems. That integration work creates opportunities for firms such as SAP, Capgemini, Sopra Steria and OVHcloud, while demand for AI infrastructure is also supporting suppliers such as Foxconn.