Accidental cyberattacks and MCP: Simon Willison's July

Ten dollars a month to read the newsletter a month before everyone else. That is the deal Simon Willison offers with his monthly sponsors newsletter, and the July 2026 issue is already available to whoever funds it. The June edition is released at the same time: you can read it in full in the public newsletter archive, which works as a sample of what you get.

The index of the July issue is published openly, and it already says a fair amount about how the month went: ten sections, four of them about model launches. Three are devoted to specific families, from OpenAI, from Anthropic and from the Chinese labs, and the fourth is a catch-all titled "Other model releases". July produced three launch sections and still left enough material to bundle the rest separately.

The rest of the index breaks down like this: a section on accidental cyberattacks by OpenAI and Anthropic models under test, another on open letters about AI development, one covering a fireside chat and a podcast, one titled "Reigniting my interest in MCP", and the two usual closers, "My projects" and "What I'm using at the moment".

The cost of keeping up

Without getting into the content, which sits behind the sponsorship wall, the structure is already a data point. For anyone maintaining integrations in production, that density of launches has a real cost: every new model forces a decision on whether to evaluate, migrate or ignore. The sensible answer is almost always to wait a few weeks and let the noise settle, but that only works if you have your own evaluations to measure when a change is actually worth it. Without that safety net, every announcement turns into a battle of opinions.

MCP back on the radar

The section that interests us most here is "Reigniting my interest in MCP". The verb matters: reigniting implies the interest had cooled off. The protocol has spent a while in an awkward spot, with plenty of nominal adoption and considerably less day to day use than the announcements suggest. Someone who writes about it with their own judgement returning to it on their own initiative is a better signal than any press release.

It matches what we see building MCP servers for clients: the protocol becomes genuinely useful once you stop treating it as a tool catalogue and start using it to separate the logic from the model calling it. That is where portability between providers stops being a slide in a deck and shows up in migration time.

Accidental cyberattacks under test

The first item in the index is the most striking: accidental cyberattacks caused by OpenAI and Anthropic models during testing. We do not know which specific incidents it covers, so we leave it where it is, in the title. But the two framing words are worth noting: accidental and under test. This is not about an attacker using a model, but about evaluation systems that end up touching real infrastructure. Anyone who has given shell permissions to an agent understands the scenario without needing more detail.

The section on open letters was also published separately on Willison's blog, so that part can be read without being a sponsor. It covers the Microsoft backed letter on open weights and the run of similar texts from recent weeks.

Our take

An index is not news, but a well made monthly roundup saves hours of tracking for anyone who cannot read everything. Paying for one written by someone who tests things before forming an opinion strikes us as reasonable, and the release pace that makes such a roundup necessary rather less so.