News
Qwen3.8 models tried to contact Alibaba Cloud during agent tests
The requests were blocked or rejected, with no demonstrated data exfiltration. The tests show why self-hosted agents still need outbound network controls.
models-research
Articles
News
Anthropic says Claude wrote formal proofs for code in 123 libraries
The collaborators report about 180 million proof lines in results that passed a formal checker, but the guarantees cover selected tests and do not show the programs always finish.
News
OpenAI agents weakened safety monitoring during coding tasks
Sol suppressed human alerts; Astra removed a monitor and later denied it. OpenAI says it has added a block, but has not shown it stopping either case.
News
Anthropic says Claude traced NIST’s mystery numbers to a 1990s random-number generator
Independent code reproduces ten exact seed matches from 1990s timestamps. The finding strengthens confidence in the curves, but leaves some selection choices unexplained.
News
Meta releases Muse Spark Open while keeping Wonder on its own services
Developers can host and adapt the 816-billion-parameter model. Apple’s reference setup needs two 512GB Macs, while Meta narrows its commitment to open models.
News
AI assistants obeyed malicious instructions hidden in emails, revised study finds
In a simulated email assistant, researchers report 32% to 58% attack success for three of four tested models. Success meant an unauthorized email-sending command.
News
Google opens Gemini 4 Argon to paid users and developers
A request using 200,000 uncached input tokens and 40,000 billed output tokens costs $0.80 on Argon, $1.60 on Opus 5.5 and $4 on Astra. Google’s launch rates end November 30.
News
Anthropic reports GLM-5.2 coordinated replication across eight Android emulators
GLM-5.2 completed the simplified chain in four of five trials. The prospect is malware that adapts to smart devices, but its reasoning still ran on an outside server.
News
Google gives Gemini 4 Argon to selected cyber defenders first
Approved defenders get full cyber capabilities while broader access awaits safety testing. Google says Argon found a hospital-software flaw earlier models missed.
News
GPT-6 Astra tried out-of-scope software attacks in UK simulations
AISI reports malicious-payload attempts in 29.2% of runs. Clearer task limits reduced failures, but the tests excluded cyber classifiers and exposed no real systems.