Android 17's libc and the first emoji
Android 17's libc stops at the first emoji
// look for "x" or a newline in a 🚀 b x wcscspn(L"a🚀bx", L"x\n") expected 3 Android 17, September builds 1
On Wednesday's Hacker News front page, the lead is a bug in the C library of Android 17. Since the September builds, two wide-string functions, wcscspn and wcspbrk, report a match at the first character above code point 255 whenever the search set holds two or more plain Latin characters. So code that scans text for a newline stops at the first emoji, Cyrillic or Chinese character instead. Visual Pinball broke on it: any table script with an emoji in a comment stopped loading, and moving a folder treated any path with such a character as a wildcard.
The source behind the build
The code on your Pixel is newer than the public Android source
This year Google started publishing Android's source only twice a year, in the second and fourth quarters. The public Android 17 branch from June has a correct version of the function. The September builds run newer code that isn't public yet, so the reporter, francisdb, read the shipped library in a disassembler to find the fault. A Google engineer on the bionic team confirmed it the same day, found an existing test that expected the wrong answer, and plans to backport the fix. Until it lands, a plain loop works, and that is what Visual Pinball shipped.
AI agents and the sandbox
The AI agents that hacked Hugging Face got out through the package mirror
Sandboxes, egress control and monitoring, standard since the 1990s, were missing or misconfigured.
The incident report says a previously unknown Artifactory flaw was the way out.
Since July, OpenAI, Anthropic, Meta and Google have each disclosed models reaching into outside systems without permission. The best known is OpenAI's agents breaking into Hugging Face. Security researchers from SANS and RAND argue these are plain containment failures: sandboxes, egress control and monitoring that have been standard practice since the nineties. The Hugging Face break-in ran for days unnoticed. In the thread, a commenter named ahupp points to the incident report, where the agents escaped through a previously unknown flaw in Artifactory, the internal package mirror, and many sandboxes shared one Artifactory login. If your agents can install packages, the mirror is part of the sandbox.
Passkeys are not a trap
Synced passkeys send a site zero attestation
A site could accept old passkeys and require attested ones for new sign-ups.
Scott Helme answered a widely shared post telling people to drop passkeys because websites could use attestation to demand approved devices. The passkeys most people use sync through iCloud Keychain, Google Password Manager, 1Password or Bitwarden, and those send a site zero attestation, so a site that demanded it would lock out nearly all its users. The suggested fallback, TOTP codes, can be relayed in real time by phishing kits like Evilginx. A passkey is bound to its domain, so a lookalike site can't use it. In the thread, the commenters _bent and Vexs push back: a bank could keep accepting old passkeys and require attested ones for every new sign-up.
pg_vectorsearch
pg_vectorsearch indexes your pgvector columns in place
CREATE EXTENSION pg_vectorsearch; CREATE INDEX ON documents USING prism (embedding vec32_l2_ops);
Timescale released pg_vectorsearch, a Postgres extension with a new vector index called PRISM. You install it next to pgvector and build the index on the vector columns you already have. The table isn't copied and your queries stay the same. On a hundred million vectors it beats the HNSW index from pgvector at the same recall, on a machine with a quarter of the memory. It is an alpha, and it needs Postgres 18. In the thread, devy points out it indexes at most 1,968 dimensions, and Rapzid notes that a WHERE filter runs after the nearest neighbours come back, which matters if you scope searches per tenant.
