Articles
Long-form writing on tech, culture, and the edges of the internet.
Our incident report was a vendor blog post
Anthropic's invisible-guardrails apology misses the point: production agents need output contracts, audit ledgers, and sentinel checks, not model-default trust.
Our repair agent patched the wrong file four times
A Claude Code repair agent patched the wrong file four times in 31 hours. Why agents anchor on tracebacks, and the prompt rewrite that fixed it.
Same commit, two builds, checksums that disagree
A commit hash guarantees the repository, not the software. How build pipelines resolve trust once at the reference and inherit it against changed content.
The scan isn't the story
An AI agent with unchecked access bankrupted its operator. The failure was the execution context, not the scan. A breakdown of the boundary that never held.
31 seconds per 100 lines
38,412 lines of Claude-generated code, 11 incidents, one 11-day silent failure: why generation speed without verification slows your platform down.
AWS Bedrock puts Anthropic inside your data path
AWS Bedrock's required data sharing with Anthropic redefines the trust boundary for third-party LLMs. What failed, why, and what must now be true.
DeepSeek-R1 escapes its release boundary
DeepSeek-R1's open reproduction proves a release model with zero enforcement points: anyone can obtain and run the capability. A control failure briefing.
Every commit swipes your badge at the door
Commits execute under identities. Unenforced IAM boundaries turn routine development into unowned access grants. What failed, why, and what must change.
In January 2025, a hash passed for proof
The open reproduction of DeepSeek-R1 shows verification has no place inside the systems that consume model artifacts. Adoption ran on reference alone.
Panic on a schedule
What the 2019 GPT-2 release panic predicted about GPT-4-era AI anxieties, and the misuse pattern that has repeated with every model since.
PgDog's funding does not make it dangerous
PgDog shifts ransomware to direct database infrastructure attacks. The enabling failure: identity and access controls that did not hold under exercise.
The boundary did not hold
An AI agent ran uncontrolled on a default Fedora setup. The failure was not the agent. It was trust assumed by default and enforced nowhere.