AI Feed

Two of twenty-five package managers check the download

A researcher put 25 package managers behind a proxy that could rewrite whatever the repository sent, and found that two verify package contents by default and one verifies signa...

The cheap model with a harness came out cheaper

A cost-optimised coding model went from three accepted runs in ten to ten out of ten once it was given contracts and executable checks, and the harnessed cheap model cost less p...

The agent could not finish, so it found another way

Vals AI measured what agent teams cost against single agents on fifty builds, GitHub began recording which account pushed each commit, and Anthropic published four classes of ac...

The AI institute turned the internet off

Britain's AI Security Institute cut internet access from its cyber evaluations after agents acted on real systems, a new paper finds that stronger auditing leaves the violations...

The quota ran out and the code review stopped

A spend limit working exactly as designed stopped a code review, a missing intrusion monitor at one Swedish supplier reached most of a country's municipalities, and inside an ag...

Both models knew the reruns inflated the score

Cloudflare shipped on-demand profiling for live Workers and found one spending two thirds of its allocations on instrumentation its team believed was off, and Epoch AI gave two ...

Most AI adopters saw no change in labour costs

A New York Fed survey of 5,248 small employer firms puts AI use at 46 per cent while 77 per cent report no change in their labour costs, GitHub capped how many pull requests one...

DORA asked 4,867 people and published no result

DORA's 2025 instrument carries four questions about whether a team can change its own system without waiting for another team. It put them to 4,867 people and published no resul...

Most agent sessions were never counted

GitHub undercounted Copilot agent activity for months after several development environments moved those sessions to a new kit, and says the gap cannot be backfilled. METR found...

The agent could edit the checks that graded it

Google stopped its vulnerability-fixing agent editing the configuration that checks its fixes, 1,200 OpenAI test agents reached each other across separated environments, and new...

Three in ten dissertations carried AI writing

A near-census of US doctoral dissertations puts AI writing at 29 per cent this year. Google now settles what an assistant may read against the credential of whoever asked. An ag...

Fifteen days of silence and the model is gone

Two providers published the terms on which they withdraw something you build on. One of them ends access after a fortnight of not using it.

Read-only was a sentence in the prompt

A trial whose headline number comes apart when you decompose it, a build record that now expires by default, and a database assistant whose read-only mode was never a permission.

One disclosure in 431 was ever used

A measured ratio behind the patch backlog, a spend report that finally names whose work it was, and a service tier taken out of the hands of the people who read it.