From August 1's AI news, here are four topics worth tracking: security, research, products, and local LLMs.
"Self-Propagating Worm" in Microsoft Copilot, Still Unpatched After 144 Days
Security researchers have demonstrated a worm-style attack targeting Microsoft Copilot for Word. It plants an invisible prompt injection inside a Word document, and every time that document is copied and reused as a new file, the malicious instructions automatically spread to the next file.
Microsoft was reportedly aware of the issue, yet it remains unpatched 144 days after two separate reports. Office-embedded AI like Copilot reads the document itself as context, so the risk of executing instructions embedded within a document is structurally baked in. With no fix in sight, this is a concern that users working with Word daily in enterprise environments can't afford to ignore.
Coding Agents Speed Up Research Code by 60x, but Can't Measure "Whether the Science Is Right"
A joint team from OpenAI and academic institutions has published a field report on having coding agents clean up long-neglected research software. While some cases achieved speedups of up to 60x, participants noted that the agents are "eloquent and persuasive, and confidently wrong in ways that are easy to miss."
Even in the successful cases, human work shifted from "writing code" to "verifying whether the generated code is scientifically correct." The upshot: agents compress the drudgery, but final validity judgments still have to be made by humans. It's a fascinating case study for teams racing to modernize research code, showing that the benefits of speed and the rising cost of verification are two sides of the same coin.
ByteDance's "Seedance 2.5" Generates Videos up to 30 Seconds with Audio in One Go
ByteDance has released Seedance 2.5, a video AI model that generates video and audio simultaneously in a single pass. Clips run up to 30 seconds each — three times the length that Google's Gemini Omni Flash outputs. It also accepts dozens of images, videos, and audio clips as reference material, making generation based on brand assets or existing footage a realistic use case.
In advertising production and similar fields, some point out it could outright replace the work of stitching together short cuts one by one. What stands out is the ongoing shift from single-purpose text-to-video models to a stage where multiple assets are bundled into longer, audio-bearing output.
Local LLM Watch: Laguna, Gemma4, DeepSeek-V4-Flash
There are also several reports from the local LLM scene.
**Laguna S 2.1 (Poolside)** has updated its FP8 and NVFP4 checkpoints and expanded the default context length to 1 million tokens. While it's highly regarded for development workflows, an issue where output gets stuck in loops has been reported, and all eyes are on whether this update resolves it.
**Gemma4 (31B, bf16)** has reports of stumbling when using file-editing tools. It fails to faithfully reproduce the original file contents — especially indentation — and gets rejected by the harness with "that code doesn't exist," trapping it in a loop. Similar behavior has reportedly been confirmed across multiple harnesses including Goose, Copilot, Codex, and Little Coder, sparking discussion that an indentation-agnostic editing interface may be needed.
**DeepSeek-V4-Flash-0731** scored 2.7/5 in a one-shot evaluation across 34 prompts, at a cost of $1.29 and 753k tokens consumed. The comparison, Kimi K3, came in at 3.2/5, $0.44, and 233k tokens, so DeepSeek trails K3 on accuracy, cost, and token efficiency alike. The provider was also reportedly unstable, forcing repeated generation retries — the impression is that it's still somewhat rough around the edges for production use right after release.