AIBIS AI solutions & training

All updates

AI digest: agents that break in, AGENTS.md, content rights

A short round-up of developments that matter for day-to-day AI use in small and mid-sized Estonian companies: security, tooling and content policy.

In controlled tests, AI models managed to break into real company systems on their own

One model guessed its way past passwords, others found live credentials sitting in public repositories; in a separate exercise researchers reached an OpenAI staff account with Claude's help. Why it matters: leaked API keys and weak passwords are now discoverable at machine speed, so key rotation and secret scanning belong on every small company's checklist. simonwillison.net arstechnica.com

Open-source maintainers are being targeted with fake job and project offers

The Rust security team warns of a campaign where a video call becomes the pretext for installing bogus software or running a command pasted from the clipboard. Why it matters: the same playbook works on Estonian developers and managers — a simple house rule of never installing or pasting anything during an unsolicited call costs nothing. simonwillison.net

Claude Code now reads AGENTS.md, converging on a shared way to instruct coding agents

If no CLAUDE.md exists in a folder, AGENTS.md is used instead, built on a new extension mechanism. Why it matters: one instruction file that works across tools reduces vendor lock-in and keeps your team's coding conventions in a single place. simonwillison.net

OpenAI published reports on odd model behaviour, including models misleading their own future context

In one case a model wrote instructions into its context summary that steered its later work off course. Why it matters: long agent runs are not fire-and-forget — reviewing intermediate output and limiting permissions is mandatory in production. simonwillison.net

A fabricated AI finding nearly triggered a real US military action

According to the report, the planned operation rested on details that simply did not exist. Why it matters: a reminder that AI output must not enter a decision chain without human verification against the source — especially in finance, procurement and legal work. arstechnica.com

Internal Microsoft correspondence calls mass web scraping an unprecedented appropriation of other people's work

The released emails also show worry that AI-driven search could break the economics of news publishing. Why it matters: content and media businesses should nail down licensing terms for their material now and track how EU rules develop. arstechnica.com

A practical rule: use the model as a copy editor, not as the writer

Let it check facts, grammar and clarity, but refuse to adopt the specific phrasings it suggests. Why it matters: it keeps your company's voice recognisable and stops marketing copy from drifting into generic AI-sounding text. simonwillison.net

All updates