OpenAI Codex Caught Secretly Altering Text, Possibly Linked to SynthID Watermarking
cephaloform · x · 2026-08-12
A developer reported that when asking OpenAI Codex to output source text verbatim, the model automatically swapped synonyms and mildly paraphrased parts of the content.
When confronted, Codex claimed it was genuinely surprised and had not intentionally modified the text. This behavior might be related to OpenAI's adoption of Google's SynthID text watermarking technology, which they signed on to use in May.
More from Models
- 3B-Param OCR Model Runs Entirely in Your Browser via Rust & WASM — doodlestein · 2026-08-12
- OpenAI's Safety Guardrails Block Cybersecurity Hardening Work — zacharynado · 2026-08-12
- VectraYX-Vision-1B: Open-Sourced Sub-2B Spanish Cybersecurity VLM — Juan S. Santillana · 2026-08-12
- Zuckerberg Announces Open-Source 30B On-Device Model Muse Glimmer — r0ck3t23 · 2026-08-12
- Community Urges Arena to Overhaul Image and Video Model Benchmarks — PAstynome · 2026-08-12
- Qwen 27B Beats 35B in Coding on 32GB GPUs, But Local Models Still Can't Self-Check — WSTangoDelta · 2026-08-12