Intezer uses AI agents to rebuild Slack, Apple Finder and a TinyTurla-NG backdoor from binaries
Intezer researchers used AI agents to rebuild working source code for Slack, Apple's Finder and a TinyTurla-NG sample, a backdoor family that Cisco Talos links to the Turla group, working only from shipped programs and vendor documentation. The rebuilt Slack client ran well enough that the team added an AI text autocomplete feature powered by a small local model. The work is a proof of concept about reverse engineering, not an active intrusion: the article reports no vulnerability, no CVE and no attack in the wild. The risk is that vendors relying on compiled binaries to protect their source code lose that barrier, and the same method can be applied to malware. Criticity is medium because there is no direct exposure, only a shift in how hard it is to recover code.
Detalhes técnicos
- Codex, running GPT-5.6-sol-high, orchestrated the workflow and made the engineering decisions, while Kimi K3 filled three separate roles. The agents produced about 3,100 lines of source in two files.
- The team set five success criteria: readable enough to maintain, clean and idiomatic, compiled only from what the agents wrote, free of bundled third-party libraries, and identical in behavior to the original. TinyTurla was rated the hardest of the three targets.
- For the TinyTurla-NG DLL, the rebuild is 296,960 bytes against 291,840 for the original, with 1,072 functions against 1,041 and identical imported symbols. BinDiff matched 1,034 original functions and scored the pair at 92.8% similarity with 98.7% confidence. The researchers say this measures structural shape only.
- Behavior was checked separately with a harness that compares each rebuilt function with its original using Binary Ninja's high-level intermediate language.
- Claude analyzed Slack's minified client to identify the third-party libraries it contains, among more than 500,000 functions. Because Slack's server is not downloadable, the agents built a Slack-compatible server on SQLite, guided by network inspection of the real app and Slack's documentation.
- The Finder rebuild, a 25.8 MB executable, still has gaps around online-only files, sync status and actions that depend on other macOS services.
- The article does not say how long the rebuilds took or what they cost.
Remediação
- Intezer advises vendors to use encryptors and obfuscators, which the researchers say "raise the bar" for AI agents. The article does not report testing how much these tools slow an agent down.
- The article reports no vendor response from Slack, Apple or Cisco, and no patch or configuration change is involved.
Linha do tempo
-
Intezer uses AI agents to rebuild Slack, Apple Finder and a TinyTurla-NG backdoor from binaries
ler a análise desta fonte ocultar
Intezer researchers used AI agents to rebuild working source code for Slack, Apple's Finder and a TinyTurla-NG sample, a backdoor family that Cisco Talos links to the Turla group, working only from shipped programs and vendor documentation. The rebuilt Slack client ran well enough that the team added an AI text autocomplete feature powered by a small local model. The work is a proof of concept about reverse engineering, not an active intrusion: the article reports no vulnerability, no CVE and no attack in the wild. The risk is that vendors relying on compiled binaries to protect their source code lose that barrier, and the same method can be applied to malware. Criticity is medium because there is no direct exposure, only a shift in how hard it is to recover code.
Technical Details
- Codex, running GPT-5.6-sol-high, orchestrated the workflow and made the engineering decisions, while Kimi K3 filled three separate roles. The agents produced about 3,100 lines of source in two files.
- The team set five success criteria: readable enough to maintain, clean and idiomatic, compiled only from what the agents wrote, free of bundled third-party libraries, and identical in behavior to the original. TinyTurla was rated the hardest of the three targets.
- For the TinyTurla-NG DLL, the rebuild is 296,960 bytes against 291,840 for the original, with 1,072 functions against 1,041 and identical imported symbols. BinDiff matched 1,034 original functions and scored the pair at 92.8% similarity with 98.7% confidence. The researchers say this measures structural shape only.
- Behavior was checked separately with a harness that compares each rebuilt function with its original using Binary Ninja's high-level intermediate language.
- Claude analyzed Slack's minified client to identify the third-party libraries it contains, among more than 500,000 functions. Because Slack's server is not downloadable, the agents built a Slack-compatible server on SQLite, guided by network inspection of the real app and Slack's documentation.
- The Finder rebuild, a 25.8 MB executable, still has gaps around online-only files, sync status and actions that depend on other macOS services.
- The article does not say how long the rebuilds took or what they cost.
Remediation
- Intezer advises vendors to use encryptors and obfuscators, which the researchers say "raise the bar" for AI agents. The article does not report testing how much these tools slow an agent down.
- The article reports no vendor response from Slack, Apple or Cisco, and no patch or configuration change is involved.