Passing tests is not preserving behavior
An arXiv paper shows that an LLM-generated decompilation can build and pass shipped tests while still changing the original program's behavior.
Desk / Signal
AI and agents
Models, agents, and the software being reorganized around them.
- Briefs
- 14
- Newest
- Sep 27, 2026
Lead brief
Desk file
- 01
Passing tests is not preserving behavior
- 02
ChatGPT Images adds a sketchbook
- 03
Adobe puts generation on the timeline
- 04
Meta gives Muse a wallet and guardrails
- 05
OpenAI puts the Codex harness behind an API
- 06
Atlassian turns Jira work into governed agent loops
- 07
Apple rolls out Siri AI across apps and screens
- 08
Anthropic folds Claude modes into one interface
- 09
OpenAI formalizes reports about model misalignment
- 10
Unity gives Codex engine-specific skills
- 11
Claude Code Projects puts agents on branches
- 12
MVLAND turns a song into a shot list
- 13
Gemini Live Avatar gives agents a face
- 14
Copilot becomes three modes