Intezer uses AI agents to rebuild Slack, Finder, and TinyTurla-NG
First merged summary.
Intezer says AI agents reconstructed working source for Slack, Finder, and TinyTurla-NG from binaries, and even extended Slack.
Intezer researchers said AI agents reconstructed working source for Slack, Apple’s Finder, and a TinyTurla-NG sample from shipped binaries and vendor documentation. Codex using GPT-5.6-sol-high made engineering decisions while Kimi K3 filled three roles; Nicole Fishbein said each target needed a different harness, and Claude reconstructed source in some cases and refused in others. The rebuilt Slack ran well enough for the team to add local-model text autocomplete, and agents built a Slack-compatible server on SQLite because Slack’s server is not downloadable. Finder still lacks some online-file, sync, and macOS-service behavior. Dark Reading reports Turla was the hardest target and recreating the proprietary apps was only moderately challenging; researchers avoided executing the malware. Intezer advises encryptors and obfuscators to raise the bar.
- Agents produced about 3,100 lines of TinyTurla-NG source in two files.
- Rebuilt DLL is 296,960 bytes versus the original 291,840.
- BinDiff matched 1,034 functions at 92.8% similarity and 98.7% confidence.
- Slack’s client contains more than 500,000 functions, mostly third-party libraries.
- Finder is a 25.8 MB executable; Intezer did not publish time or cost.
- For the malware test, researchers avoided executing anything.
What each outlet adds
beyond the shared facts- Dark ReadingTurla was hardest; Claude sometimes refused; different harnesses per target.
- Help Net SecurityBinDiff scores, line counts, Slack autocomplete, SQLite server, model roles.
- Vendors
- SlackAppleSalesforce
- Threat actors
- Turla
- Malware
- TinyTurla-NGTurla
- Organizations
- IntezerCisco Talos
- AI models
- GPT-5.6-sol-highKimi K3Claude
Coverage
oldest first · one row per article · UTCIntezer says AI eased reverse engineering of Slack, Finder, and Turla
Each target needed a different harness, and Turla was harder to reconstruct than the proprietary apps.
Intezer agents rebuild Slack, Finder and TinyTurla-NG from binaries
The TinyTurla-NG rebuild scored 92.8% BinDiff similarity, and the Slack copy gained local AI autocomplete.