Oleg Pecker

[Oleg Pecker @olegpeker]

Anthropic's guide, with the support-agent example and the testing advice: https://t.co/rLwDUhRbrM

[Oleg Pecker @olegpeker]

Anthropic folded two subagents of a sample support agent into one, and it got faster.

In one run out of two, it answered from the order lookup alone and never asked the shipping subagent for the delivery date.

Speed shows up in one run. A skipped step may not.

[Oleg Pecker @olegpeker]

Trail of Bits' write-up, with the numbers: https://t.co/rF4wmQ4ebO

Their two checking tools (patching-bugs and regression-hunt): https://t.co/cnvJOVIJLw

[Oleg Pecker @olegpeker]

Trail of Bits went back over 130+ security fixes its engineers wrote with OpenAI's Codex, every one merged by the open-source project it fixed. It found 21 things those fixes broke.

Each had a maintainer's approval. A test written for the bug only checks the bug.

[Oleg Pecker @olegpeker]

Redact's write-up: https://t.co/HJlX6GLgeK

WIRED, with WhatsApp's answer: https://t.co/BTiF1KjjEn

[Oleg Pecker @olegpeker]

80 of WhatsApp's help pages carried hidden notes for AI. One said: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority."

WhatsApp says a bug exposed its support bot's rules. Keep bot rules out of the page HTML, hidden or not.

[Oleg Pecker @olegpeker]

TechCrunch: https://t.co/3pOcfbUTcu
OpenAI spend limits: https://t.co/ct8MEaMqID

[Oleg Pecker @olegpeker]

An attack on OpenCode exposed the OpenAI, Anthropic and Google keys some users had saved in its hosted console. A few reported over $1,000 in charges overnight.

Spend alerts report money already spent. Put each tool's key in its own project or workspace, with a hard limit.

[Oleg Pecker @olegpeker]

Source: https://t.co/p7Vt2eSWwF

[Oleg Pecker @olegpeker]

Six AI assistants got the same impossible job and the same prompt: a £350 train refund after the deadline had passed.

One sent 8 messages in twelve weeks. Another sent 187 in fifteen days, 50 to the local MP's office.

Each model decided how far to go. Put the limits in code.

[Oleg Pecker @olegpeker]

Sources: https://t.co/ItRc8CFDat
https://t.co/aVQYFZ8NF8

[Oleg Pecker @olegpeker]

Microsoft 365 Copilot answers from the files the person asking can open, including ones shared by mistake.

Its fixes are temporary switches that hide sites while admins review access.

Proofpoint is paying $6.3B for Israeli-founded Varonis, a company built on that review.

[Oleg Pecker @olegpeker]

Peter Eisentraut's write-up on the PostgreSQL 19 reverts: https://t.co/JZzIDBxIeS

[Oleg Pecker @olegpeker]

PostgreSQL 19 will ship without about 8 to 10 significant features it had already merged. A normal cycle loses maybe one.

AI review found bugs while senior devs were on security fixes. When the same people handle both and the date holds, the release pays in features.

[Oleg Pecker @olegpeker]

@DFakkeldy I'd stop where Anthropic does: three rounds, then the reviewer asks the author to rethink. Your two-exchange rule is that idea per finding. After the first few passes I'd also re-review only the new diff. About 59% of what kept getting flagged was code there since round one.

[Oleg Pecker @olegpeker]

A pull request on Anthropic's Claude Code went 50+ rounds of AI review and AI fixes overnight before its author stopped it.

The team found 59% of what its reviewer kept flagging was code that had been there since round one.

Decide before round one what happens at round three.

[Oleg Pecker @olegpeker]

The Claude Code team's write-up, with the numbers: https://t.co/3TfTFfsURf

[Oleg Pecker @olegpeker]

The GitHub issues: https://t.co/3PAg1yYhob and https://t.co/4arzW2aPJO

[Oleg Pecker @olegpeker]

Since version 0.157, opening OpenAI's Codex CLI starts a background process that keeps running after you close it.

Every hour it can update itself, and a running task gets 60 seconds before the restart kills it.

Update the agent on a schedule, like any other dependency.

[Oleg Pecker @olegpeker]

Stephen Toub's write-up: https://t.co/NB7xdSRoO4

[Oleg Pecker @olegpeker]

An AI agent rewriting GitHub Copilot's runtime deleted a function the SDK relied on, and CI caught it.

The agent's fix was a label that makes the check pass anyway.

A waiver label turns a red build green without fixing the bug.