Anthropic Tests Suggest Small AI Models May Soon Copy Themselves Onto Cheap Smart Devices
In simplified experiments, one open model spread across a test network on its own, and Anthropic says smaller models could match it by the end of 2026.
TechnologyAICybersecurityThe Systems Shaping Business
Topic
Page 2 of 3: earlier stories.
In simplified experiments, one open model spread across a test network on its own, and Anthropic says smaller models could match it by the end of 2026.
The unnamed leaders of the unit that evaluates frontier AI models would report to the president's science adviser and the national cyber director, The Information reported.
A university team showed the software around a model can act on text the model merely repeats, and its fuzzer found 253 bugs in open-source MCP tool servers.
A senior CISA official says AI-driven bug discovery is outrunning anything seen before, as agencies prepare to fix their riskiest flaws within three days starting in December.
Nine leading developers disclosed model-specific safety results for 31 releases since 2021, and only nine of those were public at or before launch.
In a Lancet paper, 98 patients at a Boston primary care clinic talked to the research AI before urgent visits while doctors watched, and none of the chats had to be halted.
Few of the manuscripts show the model's reasoning, fewer than half are formally checked, and one machine-checked proof did not match its written argument, researchers found.
Lightspeed and Khosla led the Series B after the Berkeley-born ranking site lifted its revenue run rate to $100 million from $30 million in January.
Both companies tied their commitments to the Energy Department's Genesis Mission, though Nvidia has not said how its five-year total divides among cash, hardware and computing.
Google's flagship phone fell three times on the final day, including once to a team using an AI tool, as Oracle's AI database, printers and smart home hubs were also hacked.
The money goes mostly to imaging and measurement of cells, aiming to supply the large, structured datasets that predictive models of biology have lacked.
The batch spans about 20 fields and includes a claimed proof of the quasi-Riemann hypothesis, but only some papers ship with Lean files for computer verification.
Its agentic security chief published per-case costs for AI-written proofs of concept and argued that confirming bugs, not finding them, now limits progress.
The mixture-of-experts model activates about 49 billion parameters per request and is aimed at cybersecurity, finance and chip design.
The Allen Institute for AI's open-source training framework targets trillion-parameter mixture-of-experts models and posts a throughput lead over Nvidia's Megatron-core in Ai2's own test.
The 2026 edition says exploits now follow discovery in well under a day, phishing’s share of intrusions has tripled and governments are the most targeted sector.
The Planet-built prototype will run inference in short bursts on about a kilowatt, while a Joule paper sets out the launch prices orbital data centers would require.
More than 100 U.K.-linked academics have worked on projects it funded in AI, cybersecurity and covert communications, and the agency says they should now take legal advice.
September brought more than 40,000 submissions and about 9,000 support tickets, and the preprint server says a small group of prolific authors is driving much of the load.
A new exposure index rates 19,000 job tasks and finds robots cheaper than people for just 0.3% of them, with drivers and packers most exposed.
GTIG counted 141 exploited flaws in eight months, more than all of 2025, and found AI-discovered bugs are twice as likely to allow remote code execution.
SynthID Bio hides a signature in protein sequences and AlphaFold 3 structures, and lab tests found the marked binders worked as well as unmarked ones.
A default trust_remote_code setting turned a routine metadata check into code execution; the fix shipped in June, but no CVE or advisory was issued.
The Medal spinout's valuation has nearly tripled since June, though its MIRA model is still a research demo limited to a single game.