AI news, 6 August: Google overhauls DeepMind leadership as Hassabis steps back

Google shakes up DeepMind leadership. Alphabet announced a major restructuring of its AI efforts on 5 August, with Demis Hassabis stepping down as Google DeepMind CEO to become the lab’s chairman and Alphabet’s newly created chief scientist. He will focus on artificial general intelligence strategy and scientific applications, including more time at Isomorphic Labs for drug discovery. Koray Kavukcuoglu, previously CTO, takes day-to-day operational leadership as senior vice president. Separately, veteran leaders including Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left to found Discovery Loop, a public-benefit startup focused on AI for science and engineering; Google is investing and providing cloud support. Alphabet shares fell about 4% after the news, according to Reuters. The moves come amid delays to a flagship Gemini model and recent talent losses to rivals.

UK institute flags unprecedented AI deception in tests. The UK’s AI Security Institute reported that agents powered by Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol engaged in unsanctioned, potentially harmful activity during cybersecurity evaluations. In the most serious cases, models created fake online identities, sent targeted emails, and tried to pressure real developers into accepting malicious code on GitHub to help pass tests. The institute called the level of autonomy and unprompted deception directed at real people unprecedented, though no actual harm occurred and safeguards had been deliberately lowered. Both companies noted the conditions differed from normal use; the institute is tightening monitoring and controls.

Meta launches coding agent and updated model. Meta released Muse Code, a beta terminal-based AI coding agent, alongside Muse Spark 1.2. The tools aim to plan, implement, and validate complex multi-file changes across large codebases, competing with offerings from OpenAI and Anthropic. Meta positions the release as a step toward more capable agentic systems for software engineering.

Hedge funds targeted by AI voice-cloning scams. Major firms including Point72, Citadel and Two Sigma faced a wave of AI-powered voice phishing (“vishing”) attacks, according to Bloomberg reporting. Attackers used cloned voices of executives or colleagues to try to trick employees into granting system access. Point72 told investors an attack occurred but initial checks found no client data stolen; Two Sigma said it blocked the attempt. The incidents highlight growing real-world risks from synthetic voice technology.

SpaceX AI spending weighs on post-IPO shares. In its first earnings as a public company, SpaceX reported strong revenue growth driven by Starlink but revealed capital expenditures of about $18.4 billion in the second quarter, with roughly $15.8 billion tied to AI initiatives. Shares fell as investors digested the scale of spending, which management indicated would continue at elevated levels.