Today's issue
The agents went to work — shipping pull requests from a chat box, reading a thousand codebases for flaws, auditing the spreadsheet you forgot to check. Every one of them is a new thing to defend.
Ephemeris.
Someone is reselling your model calls.
Inference theft is the quiet line item nobody budgets for: an attacker funnels your paid API quota into a product they sell as their own, and you find out from the invoice. Vercel's piece explains why rate limits and API keys miss it — the abuse looks like ordinary traffic — and makes the case for catching it at the request layer with BotID instead of after the bill lands.
Ship code straight from Slack.
PostHog wired an agent into Slack: mention it with a feature request and it spins up a sandbox, writes the code, runs the tests, and opens a pull request a human still reviews. A two-stage classifier routes each ask. Over two weeks it merged 116 contributions — a concrete look at what happens when the people closest to the problem can file a fix, not just a ticket.
An AI read the code and found ten thousand ways in.
Project Glasswing hands select organisations a security-tuned model, Claude Mythos, that scans their codebases for vulnerabilities. Early partners surfaced more than 10,000 high- and critical-severity flaws. This expansion widens access from roughly 50 to about 150 organisations across 15-plus countries, weighted toward critical infrastructure — power, water, healthcare, communications — where an unpatched flaw is more than a bug.
Codex opens up beyond engineers.
OpenAI gave Codex three extensibility hooks — plugins, sites, and annotations — pitched not only at developers but at analysts, marketers, designers, and investors. The bet is that the agent is more useful when it plugs into the tools a team already runs and can publish or annotate what it builds, rather than living behind a terminal.
Plugins
Connect Codex to the tools and data a team already uses.
Sites
Publish what the agent builds as a shareable page.
Annotations
Leave inline notes the agent reads and acts on.
Put the numbers where the decision happens.
Analytics users never see still drive the product they pay for. PostHog catalogues five ways teams expose data to customers and argues the craft is placement and trust, not dashboards.
The pattern that fails is the one buried in a separate reports tab. The ones that work sit inline, at the moment a customer is deciding something — and they hold to sub-minute latency, because numbers people don't trust they stop reading.
Treat customer-facing analytics as a core feature with an owner, the piece argues, not a side project the data team squeezes in. Scope every figure to the viewer who's allowed to see it.
Five patterns in the wild
- Embedded dashboard
- Vercel
- Inline metric at the decision
- Booking.com
- Public leaderboard
- Product Hunt
- Gamified contribution graph
- GitHub
- Augmented internal data
- in-product
A second reader for the spreadsheet you trust.
excel-hygiene is a small Claude skill that audits an .xlsx file two ways: a Python pass catches the mechanical errors — numbers stored as text, formulas overwritten with static values, truncated sums, percentages off by 100× — while the agent reasons over the Panko–Halverson error taxonomy for the logic mistakes a machine check misses. Clone it into your skills folder and ask it to look.
| A | B | C | |
|---|---|---|---|
| 1 | Revenue | =SUM(B2:B9) | ✓ |
| 2 | Q1 | "12000" ⚑ text | |
| 3 | Q2 | 13400 | |
| 4 | Margin | 0.62 ⚑ formula → static | |
| 5 | Growth | 8 ⚑ pct as whole |
check_excel.py flags the mechanical errors · the agent reasons about the rest
That's today.
Six picks on a thin day — the thread running through them is agents doing real work, and the new surfaces that work leaves exposed.