30 years shipping code. Full-stack developer, AI-tool user and AI-hype sceptic. Writing about craft, security, and what nobody's saying about agents.
Joined August 2014
- Tweets 57
- Following 25
- Followers 25
- Likes 19
Photos and videos
Jun 13
Every resilience playbook we have, retries, multi-region, fallback, assumes the provider wants to stay up. None of it survives a government order.
How are you planning for an AI dependency that can be made illegal, not just unavailable?
#AI #SoftwareArchitecture
8
Jun 13
A US export order pulled Claude's strongest model worldwide on June 12. The "jailbreak" behind it? Asking it to read a codebase and fix bugs. Your day job got a model recalled. You can't retry your way out of a subpoena.
tim-schipper.nl/en/blog/the-…
#AI #ClaudeCode #AIAgents
45
Jun 10
Ok, I have 3 accounts .. 2 for personal use; a Gemini AI Pro and Claude Pro and for work a Claude Max.
Just ran `npx ccusage@latest daily` on both my dev boxes:
Box 1: $5644.35
Box 2: $5157.91
Tic toc ... how long till the bubble goes plop ?!
1
60
Jun 10
Genuine question for anyone running frontier models on real work:
your model flags a legitimate task as "too dangerous" and downgrades you to a weaker one. It does tell you it happened.
Is a polite downgrade still a broken tool? Where's your line?
11
Jun 10
Claude Fable 5 flagged my security audit as "too dangerous," handed it to a weaker model, then admitted in writing that it flags safe work too.
The most powerful public model ever shipped. A Ferrari with a 30mph limiter.
tim-schipper.nl/en/blog/the-…
#Fable #AI #AiAgents #ClaudeCode
33
Jun 9
On 10 May an LLM agent broke into a database on its own. A known CVE to full exfiltration, four pivots, under an hour. No human typed a single command.
The tell wasn't skill. It was speed: 12 API calls across 11 IPs in 22 seconds.
1
31
Jun 9
And now you're installing one yourself.
At Build 2026 Microsoft put an agent in Windows with inbox access, on by default. One that reads your email and acts on your files is the perfect prompt-injection target. The attacker just sends you an email.
tim-schipper.nl/en/blog/the-…
1
12
Jun 9
Nobody's driving the attacker. Nobody's driving the agent on your own desktop either.
The "frictionless, on by default" pitch quietly removed a safety feature.
#AI #AIAgents #CyberSecurity
11
Jun 9
2023: "AI will replace developers."
2024: "AI will make developers 10x faster."
2025: "AI will write the code."
2026: Developer spends 3 hours debugging code written by 4 different AI agents arguing with each other in JSON.
Nature is healing.
10
Jun 6
Your coding agent has no world model. You built it one.
The compiler, the type checker, the test suite that goes red, the loop that feeds errors back in. That harness is a world model sitting outside the weights instead of inside them.
Raw ChatGPT was a party trick. Same model wired to a toolchain is a colleague.
tim-schipper.nl/en/blog/your…
#AI #AIAgents #ClaudeCode
1
24
Jun 5
My agent spent an afternoon fighting a bug I'd already fixed. It "remembered" the old failure and waved real ones through.
Persistent memory isn't a pure win. It stores a moment, not a truth, and moments go stale.
tim-schipper.nl/en/blog/stal…
#AIAgents #ClaudeCode
1
65
Jun 5
Your agent writes itself a note that's true today. Next week the code changed and the note didn't. Now it's confidently wrong.
How often do you actually read back what your coding agent has "remembered"? And do you prune it, or just let it pile up?
24
Jun 5
49 subagents in parallel. One session. An estimated $8k-15k bill.
Dynamic Workflows made speed free. They didn't make judgement free. A thousand agents just generate the reviewing backlog faster than you can clear it.
tim-schipper.nl/en/blog/spee…
#AIAgents #ClaudeCode #LLMOps
1
1
33
Jun 3
GitHub Copilot costs are up 10-50x for some devs. Everyone's angry at GitHub.
Wrong target. The bill isn't the surprise. The surprise is finding out what your workflow was actually worth all along.
tim-schipper.nl/en/blog/the-…
#GitHubCopilot #DevTools #AITooling
1
159
Jun 2
Hackers took over the Obama White House Instagram this weekend by asking Meta's AI support bot to reset the password. No exploit. No breach. They didn't even need the victim's email.
tim-schipper.nl/en/blog/they…
#AIAgents #AppSec
196
Jun 1
Uber burned its entire 2026 AI budget in 4 months. Their CTO: "no link yet between AI tokenmaxxing and shipping products."
Amazon built a token-usage leaderboard. Employees gamed it. Obviously.
Goodhart's Law in action.
tim-schipper.nl/en/blog/toke…
#GoodhartsLaw #AIAdoption
1
1
28
May 29
Opus 4.8 beats GPT-5.5 and Gemini on SWE-bench. That's the least interesting thing about it.
Real upgrade: 4x fewer unflagged flaws in its own code. A model that knows when it's guessing beats points of benchmark every time.
tim-schipper.nl/en/blog/the-…
#AICoding #ClaudeCode
1
58
May 28
An attacker wrote malware to rob Claude's workspace, then leaked his own GitHub token inside it.
Both sides vibe-code now. Your agent is the target: your keys, your installs, your configs.
tim-schipper.nl/en/blog/even…
#AIAgents #AppSec
88