Google open-sources HEIR: AI inference on encrypted data
Compile a pretrained model once and serve predictions on data you never decrypt — four production demos included. Plus: Claude Code auto mode is now default.
Copy markdown
Compile once, compute on ciphertext
HEIR converts a pretrained model into a version that runs inference on encrypted inputs — the server returns predictions without ever seeing the plaintext. Google frames the target as one-click encrypted inference for developers who are not cryptographers.
Four demos you can actually copy
It ships production-shaped examples: a deep-learning recommender, credit-card fraud detection, network-intrusion detection (Kitsune), and audio hotword spotting. If your product touches data users will not send in the clear — health, finance, voice — that is your weekend build.
The catch: privacy now, speed later
Latency is quoted on a single-threaded CPU, so this buys confidentiality, not throughput, today. Google has named accelerator partners (Belfort, Niobium, Cornami, Optalysys) to close the gap, and four peer-reviewed papers already build on the stack.
Elsewhere: Claude Code stops asking first
Auto mode is now the default on Pro, Max, and Team plans: tool calls route through a classifier that blocks irreversible or out-of-environment actions instead of prompting you each time. Anthropic says it caught 89% of dangerous commands in testing versus 13.6% for human review; flip it back with Shift+Tab or pin disableAutoMode.
Elsewhere: ChatGPT desktop lands on Linux
OpenAI’s Linux desktop app is in public preview (Ubuntu, Debian, Fedora) and ships with Codex — a native option for devs who had been stuck in the browser. Early reviews are mixed, but it is a real install, not a web wrapper.