AI agents have rebuilt working versions of Slack, Apple’s Finder and a malware sample from their shipped software, according to research by Intezer. The finding has a useful double edge: it shows how AI might help recover software, and how the same technique could expose code that vendors keep private.
Watch Desk analysis
What happened
Help Net Security reports that Intezer researchers used AI agents, with Codex running GPT-5.6-sol-high directing the work, to reconstruct the three targets from software packages and vendor documentation. For TinyTurla-NG, a backdoor family that Cisco Talos attributes to the Turla group, the researchers produced about 3,100 lines of source code. They say the rebuilt version behaved identically to the original under their tests; a binary comparison tool matched 1,034 original functions and scored the pair at 92.8 per cent similarity.
The targets were not all complete replicas. Slack’s server was unavailable, so the team built a compatible server using SQLite and information from the app’s network traffic and documentation. The Finder reconstruction still lacked some online-file, sync-status and macOS-service functions. Intezer did not report how long the work took or what it cost. Its research account, as reported by Help Net Security, describes the results and the limits.
Why it matters
Reconstructing source code from a shipped program could help researchers understand software when its original code is unavailable. It also complicates the idea that compiled software stays secret simply because the source was never published. The malware example makes the security implications particularly clear, although this report describes a research demonstration, not evidence of a new attack in the wild.
Our read
This is a meaningful demonstration, not proof that AI can neatly reproduce any application at low cost. The missing Finder features and unavailable Slack server are important boundaries, and the researchers did not publish time or cost figures. Even so, reconstructing and testing substantial code from binaries is more than a clever autocomplete trick. Software vendors may need to treat obfuscation as a way to raise the difficulty, not a force field. That distinction is less comforting, but more useful.
What to watch
- Whether Intezer publishes further detail on the time, cost and repeatability of its reconstructions.
- How independent researchers assess the reported tests and behavioural matches.
- Whether vendors change how they protect software whose source code is not public.
Discussion spark: If AI can make compiled software easier to reconstruct, should vendors prioritise stronger obfuscation, or accept that shipped code is ultimately inspectable?
Sources and evidence
- Source update (10 November 2026, 14:45 UTC)
- Introducing Kimi K3.1 Pro on Amazon Bedrock (10 November 2026, 14:45 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.