trust-boundaries
13 posts
An open door where the gate should be
GitHub's AI agent returned private repo content when tricked, proving it holds read reach across the private boundary with no enforced refusal.
Your subject can end your investigation
A US ambassador used Belgian police to halt reporting. The failure is an unscoped trust channel from subject to enforcement, not a press dispute.
Trusted is a label, not a boundary
US authorization of Mythos AI grants access by a 'trusted' label with no confirmed revalidation, monitoring, or revocation. A label is not a control.
Saying you built it proves nothing
A contested 'vibe code' claim shows why self-reported origin accepted without verification is an unenforced control, not a trust boundary.
They walked out with the blueprints, not answers
Anthropic alleges Alibaba extracted Claude capabilities. The confirmed issue is structural: authenticated access governs entry, not what a party accumulates.
A bypassed control is worse than no control
CSSQuake is a demonstrable bypass of web application protections through trust manipulation and control boundary erosion. A breakdown of what failed and why.
No instances, bigger targets
ATProto has no instances, but the trust boundaries didn't vanish - they consolidated into plc.directory, DNS handle resolution, and rotation-key custody.
The door was unlocked, not picked
Federal concern over fable 5 was a trust boundary failure, not a jailbreak. Fix this code targets content, not access enforcement.
A broken boot is not a breach
macOS 27 beta stops Asahi Linux booting. Not a confirmed breach: a demonstration that the assumed boundary between OS domains was never enforced.
Boundary change, not version bump
WASI 0.3 shifts IPC to shared memory. If runtime isolation is not enforced, your trust boundary does not exist. An operator briefing on what must change.
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.
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.