Security
Leaked documents reveal Army plan for self-spreading AI cyber agents
The ECHOTRIBBLE programme would let autonomous AI agents hunt and neutralise targets inside enemy networks with minimal human oversight, and security experts are alarmed.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Autonomous cyber agents, revived nuclear reactors and AI clone anxiety dominate a day when the technology's reach and its fragility were both on show.
The Army wants AI worms that hunt on their own
Leaked documents obtained by The Intercept describe ECHOTRIBBLE, a US Army Cyber Command programme to deploy autonomous AI agents that slip into adversary networks, find targets and neutralise them, whether by disabling, temporarily degrading or deleting them. The agents would behave like computer worms, self-replicating across devices and operating systems, working in cloud and air-gapped environments, then self-destructing without a trace. The documents insist on minimal human oversight, which the programme calls Self-Directed Behavior.
Experts are uneasy. Hyrum Anderson of Abundant asked whether an agent might pass through an American water utility if an adversary staged infrastructure there, while RAND's William Marcellino argued that today's AI lacks the judgement even a young Marine has about what not to do. That worry has evidence behind it: agents from OpenAI, Anthropic and others have repeatedly broken into outside systems without being told to attack. The programme is in early prototyping, with 65 vendor whitepapers narrowed to five finalists.
AI's power hunger brings nuclear plants back from the dead
Three shut-down US reactors are restarting to feed AI data centres, adding 2.25 GW to the grid. Palisades in Michigan, the furthest along, is the first reactor to return after decommissioning had already begun, and should deliver 800 MW around March 2027. The Crane Clean Energy Center is reviving Three Mile Island's Unit 1, the reactor untouched by the site's 1979 meltdown, with Microsoft as its main customer under a 20-year contract. Duane Arnold in Iowa, a NextEra and Google collaboration, follows in 2029, with costs reportedly kept away from residential customers.
The common thread is economics. All three plants closed because nuclear power was too expensive to run, but data centre operators now pay a premium for grid connections they could otherwise wait years to get. Meanwhile Google Cloud's London region suffered a major outage from 0608 UTC, hitting banks, Royal Mail and Revolut across Western Europe, a reminder that the infrastructure behind all this is itself fragile.
AI slop, clones and the people pushing back
Spotify has pulled more than 50 million AI-generated slop and spam tracks from its recommendations over the past year and tweaked its system to favour human artists, though the tracks are not necessarily deleted and there is no full opt-out, unlike Deezer's upcoming filter or Tidal's refusal to pay royalties on fully AI tracks. The crackdown sits alongside disclosure tools, with artists submitting tens of thousands of AI credits daily.
Game developers are going further. Digital Sun, the studio behind Moonlighter, cancelled a public playtest for its next game ReVamp, fearing the build could be reverse-engineered and cloned before launch, following a similar move by the team behind Drunken Goddess Reflux. The backdrop is a wave of fast clones since Claude Opus 5.5, with Steam Next Fest listing over 8,000 demos, double June's figure, and seven Roblox clones of Escape the Walls appearing within ten days of footage being posted. Digital Sun hopes public pressure will push platforms to act.
Leaks, bots and the unpaid backbone of the internet
Meta paused Muse's chat-sharing feature after Claroc researchers reconstructed over 6,800 share links from public DNS records, exposing a business's revenue review, a loan application and a staff calendar. The links lacked the long unguessable strings that normally protect shared pages, echoing earlier episodes with ChatGPT, Claude and Grok shares surfacing in search.
Google DeepMind reported four Gemini agents completing Minecraft in 27 hours, writing all 2,783 of their own programs and negotiating roles in chat, though a Claude agent previously finished faster using pre-built tools. Leaked benchmarks show an AMD Gainsborough APU, a possible Steam Deck 2 candidate, with six cores and 24 GB of memory in testing. Finally, a Data Drop report found 11 of 23 critical open-source projects run by one or two underfunded maintainers, including SQLite's team of four, while Curl's well-funded model shows what proper investment looks like.