Blog
Silence You Can Trust — A One-Person Alerting Setup, Built by an Agent
On September 28, after a Docker upgrade, the reverse proxy on one of my servers went into a restart loop. CPU, memory and disk were all fine. Anyone looking only at resource graphs would have seen an uneventful day. What made my phone buzz was the HTTP check sitting right next to them. Build any automation that sends alerts and you hit the same problem quickly. “Nothing to report today” and “the thing that reports is dead” are equally quiet. A morning with nothing on your phone might be a calm morning, or a morning where the watching stopped. An alerting system is only useful once that silence can be trusted.
2026-09-29
Only the User Can't Read It
Two API calls are enough. First, ask a capable model anything. The answer comes back with a long, unreadable string attached — the model’s private notes, sealed. On the second call, copy that string into a request to a cheaper model from the same provider and tell it to transcribe the contents. The cheap model complies. What stands out is that the capable model was never broken. It does not even appear in the second call. The seal was never cracked either. Yet the contents came out.
2026-09-14
When the Brain Grew, the Problem Wasn't Search. It Was Tokens.
I was watching five whole documents get poured into a context window to answer a single question. This was the path an agent took through our team’s shared brain: pull search terms out of the question, grep the markdown files, open the top five by match count, and read until it hits the answer. I modeled that path mechanically and measured it. An average of 9,391 tokens per question. The answer was usually in there. The problem wasn’t whether the answer existed — it was that pulling out one line cost five documents’ worth of reading.
2026-09-09
I Skipped the Handover
When someone new joined the team, I gave them no handover. I did not sit them down and walk through the shape of our systems, and I did not tell them what to install first. I handed over a link to an onboarding page, and that was all. Not long after, their development environment was standing. Credentials were in place, the internal services were connected, and they were running the team’s working workflow as it is. Except that the person did not perform that setup. They handed the onboarding page to an AI agent, said ‘set this up as written’, and the agent executed it from beginning to end.
2026-08-25
While We Waited for Quantum Computers, AI Arrived First
“AI just broke post-quantum cryptography.” That headline swept through the security community in late July, following an announcement from Anthropic that its model had found a key-recovery attack against HAWK-256, a NIST post-quantum signature candidate. The grammar of the reaction was familiar. The old dread — quantum computers will bring down every cipher at once — had returned, with only the subject swapped out for AI. In a previous post, I dismantled the quantum panic by classification: cryptography is not one monolith, and quantum computers topple only one side of it. The same tool applies here. Don’t stop at the verb ‘broke’ — separate precisely what broke from what did not.
2026-07-29
Is This My Writing?
A number from lcamtuf made the rounds, summing up Hacker News in a single line. In June 2026, roughly 55% of the daily top stories were either about AI or written by AI. Four months earlier, in February, it was 40%; now it had crossed the halfway mark. His method was simple. He classified the top stories by hand each day, ran the suspected ones through a detector, and verified the flags by eye.
2026-06-25
Proving Provenance Doesn't Stop Malicious Code
Lately the security world abroad has been busy over a single incident. On May 19, 2026, more than 300 malicious versions were published across 323 packages on the npm registry in the span of 22 minutes, reportedly. The key point is that no human did this. There was no attacker sitting at a keyboard infecting packages one by one. The code itself stole maintainer tokens, used those tokens to replicate itself into other packages, and stole more tokens from there — all without human intervention, the reports say. A self-replicating worm called “Shai-Hulud” is said to have spread through the npm ecosystem this way.
2026-06-18
Documents as a Neural Network: I Put a Ruler to My Own 'Second Brain'
Anyone who opens Obsidian’s graph view for the first time tends to freeze at the same picture: hundreds of dots tangled in lines, slowly pulsing, like a cross-section of a firing brain. It’s the exact screenshot every “second brain” pitch puts on its first slide. The promise: notes accumulate, they wire up like synapses, and at some point the system starts thinking for you. I believed that promise and piled up notes for nearly a year. A year later my graph looked like a brain but behaved like a junk drawer. Plenty of dots, most of them isolated islands touching nothing, and the documents I actually needed I was finding through global search, not the graph. Between the slogan — “turn your documents into a neural network” — and what I actually held in my hands, there was a clear gap.
2026-06-16
Quantum Computers Won't Crack Your Encrypted Database
“Once quantum computers arrive, won’t your database encryption solution be useless too?” When you tell people you build security products, this is the question you hear most often these days. A vague dread sits behind it. The moment the word quantum shows up, an apocalyptic image follows — every lock in the world springing open at once. There is one decisive misconception buried in that fear: it treats “encryption” as a single block. In reality, encryption splits into two camps with entirely different characters, and a quantum computer topples only one of them. The side that falls is not the encryption that locks data itself, but the encryption that establishes trust when data is exchanged. Fail to draw that line, and you will forever misjudge the size of the threat.
2026-06-13
Fable 5: Warp Drive, or Marketing Mirage?
I ran Claude Code with Fable 5 and requested security analysis across several rounds, hoping to verify the ‘Mythos-class’ security performance Anthropic has been touting. The result was a letdown. The analysis kept getting bounced back to Opus 4.8, and on the rare run where Fable 5 carried the work through to the end, the output never stepped an inch beyond the ground Opus had already covered. When the Guardrail Blocks the Defender Anthropic’s line is that it blocks abuse by attackers while permitting requests from defenders. In practice, though, Fable 5 kept rerouting even defensive vulnerability-analysis requests to Opus. My attempt to outpace attackers with a top-tier agent was braked, again and again, by the conservative guardrail the model imposes on itself. Can we really call it ‘safety’ when the security tool ties the defender’s hands?
2026-06-11