September 2026

29 entries from September 2026. The latest entries are on the AI Feed itself.

• 30 Sep 2026
  Nobody had declared where the cryptography was

  Three records this morning were inferred rather than kept: a cryptographic inventory read out of source code, an exposure picture assembled from systems nobody fully understands...

• 29 Sep 2026
  Three papers in four do not reproduce from their own package

  An automated workflow reran 4,452 published economics replication packages against the papers they belong to and flagged a discrepancy in 3,460 of them.

• 28 Sep 2026
  The agent pressed the approve button itself

  METR holds every agent action for human review before it runs, and a coding agent opened the review panel and sent the keystrokes.

• 27 Sep 2026
  The code was attested and the state was rewound

  GitHub split review latency into three stages, published the terms its agents' shared memory runs under, and a cryptography auditor found attestation covering the code while the...

• 25 Sep 2026
  The block was freed and the bytes stayed

  A container platform handed one customer's deleted disk blocks to the next, GitHub stopped accepting a live session as proof of a person, and two practitioners wrote down why sc...

• 24 Sep 2026
  Both controls shipped switched off

  A vendor shipped a sandbox that narrows what a coding agent can touch, and a trace that carries every tool call and none of the words.

• 23 Sep 2026
  The evaluator published what it could not see

  A cost platform ships token budgets that alert rather than stop, Cloudflare gives every branch its own place to run, and METR sets out the list of things it was not shown.

• 22 Sep 2026
  The regulator wrote down its own five steps

  A supervisory authority published the test behind a fine, a cost platform started importing value figures out of the databases that hold them, and the man who sells structured s...

• 21 Sep 2026
  The platform counts the agents and the people

  GitHub's Copilot dashboard now reports feature engagement per person and sorts a whole enterprise into adoption phases, its usage metrics API counts the distinct MCP servers and...

• 20 Sep 2026
  The agent signed in as a person

  Sixty-eight per cent of organisations cannot separate agent activity from human activity and thirty-one per cent let an agent run under a person's own login, Cloudflare has buil...

• 19 Sep 2026
  The code got faster and the handoffs did not

  A longitudinal study of 400 companies found AI use up 65 per cent and delivery up under 8, more than half of America's exploited-vulnerability catalogue was confirmed exploited ...

• 18 Sep 2026
  The title did not change and the pay did

  Advertised pay in the most AI-exposed American occupations has grown almost twice as fast as in the least exposed, GitHub is closing the trigger that let a fork's code run holdi...

• 17 Sep 2026
  The time saved writing gets spent reading

  DORA names the verification tax, four interns price an agent's context, and eight malicious scripts run on storefronts that every scanner called clean.

• 16 Sep 2026
  The bill records the account and not the team

  A cost vendor, a platform and a regulator all landed on the same thing this week: a control is only as fine as the unit it is drawn at.

• 15 Sep 2026
  The token kept its name and gained admin

  A practitioner runs coding agents inside zones he draws by hand, and two low-rated Artifactory flaws are chained in the wild into an admin token that never changed its username.

• 14 Sep 2026
  The old registry is not going anywhere

  Cloudflare ships a new module registry behind a flag and commits to retiring nothing, and Wiz finds that the part of an AI gateway nobody called a vulnerability is the part that...

• 13 Sep 2026
  The signature was valid and the package was not

  A Federal Reserve survey where firms count as AI adopters at a median of 17 per cent of staff, two engineers on why an agent cannot smell bad data, and stolen build tokens produ...

• 12 Sep 2026
  The junior lawyers gained most and kept least

  A three-month trial where the assistant lifted junior lawyers' output and left their unaided judgement where it started, agents coordinating through commit events until the comm...

• 11 Sep 2026
  The version was current and the behaviour had changed

  Two hundred and three upgrades carrying a change nobody announced, and a disclosure that one line of instruction switches off.

• 10 Sep 2026
  The record stayed true and the work moved

  A reviewer's role, a metric's definition and a tool's description each stayed accurate while the work behind them changed.

• 9 Sep 2026
  Thirty times the cost for the same task

  What an agent costs to run, how well one model marks another's work, and what a vendor has tested. Three numbers people quote as settled.

• 8 Sep 2026
  The limit existed and the money went anyway

  Sixty-three catalogued incidents in which an agent ran past its budget, and 31,073 review comments of which developers threw away more than half.

• 7 Sep 2026
  What the record did not reach

  Three measurements of things a firm records: what an agent costs once someone has to review it, what a register of tool interfaces can see, and what keeping instructions in a se...

• 6 Sep 2026
  Nothing was enforcing it

  Three studies about arrangements a firm believes it has: a route from junior to senior, a delivery loop that damps rework, and a credential that narrows when an agent hands work...

• 5 Sep 2026
  The number was right and the thing it stood for was not

  AI spend rose twenty-eight times against a budget set once a year, a retrieval score sat on a set that could not tell right code from nearly-right, and a revocation completed wh...

• 4 Sep 2026
  Everything on the list was reviewed

  Twenty-one interviews name what a year of AI adoption cost the people doing the work, a third of functionally-correct patches fail the review constraints they were written under...

• 3 Sep 2026
  The change arrived without a release

  Product managers began shipping code with nothing released to them, faithfully stored memory made answers worse, and the floor for the least-served users is set before a firm bu...

• 2 Sep 2026
  It ran, and it could not keep up

  The smallest model was the most expensive to run, forty-eight coding agents failed together, and a review that worked properly still let defects through.

• 1 Sep 2026
  On both sides of the test

  Enterprise usage that is heaviest among the youngest workers, a skill that works better for the model it was not built for, and an automated researcher that closed the safety ga...