Intezer says AI eases reverse engineering of apps and malware, Dark Reading
Dark Reading reports Intezer researchers say AI eased reverse engineering of Slack, Finder and Turla; vendors may want to review what code they ship.
- Published
- Exploitation
- No confirmed exploitation of a vulnerability; Dark Reading relays an Anthropic report on threat actors using Claude for reverse engineering, not independently verified
What happened
According to Dark Reading, researchers from Intezer say mainstream software and malware can be reverse-engineered more easily with AI tools than before. This is a single-outlet report on research findings. We did not see the underlying Intezer write-up, and we have not independently verified the claims.
Dark Reading reports that Intezer tested AI-assisted reverse engineering on Slack, Apple’s Finder app and the Turla malware. Per the article:
- Intezer senior security researcher Nicole Fishbein says AI “lowers the bar”, because reverse engineering used to require experience and a lot of time.
- The researchers needed a different “harness” for each target. Their malware harness was a deterministic Python control plane around a compiler, with a virtual machine, static analysis tools and multiple LLMs. Dark Reading says an orchestrator agent split the binary into functions using tools such as Binary Ninja, and one agent reconstructed code while another tested it.
- Dark Reading reports they built a Turla replicant with its full working pipeline, including command handling and file transfer. The malware was the most difficult of the three targets.
- Finder was recreated only partly: the copy supported everyday uses, but its internal structure could not be mimicked precisely.
- Slack’s client was reproduced “very simply” by reverse-engineering its JavaScript source, but the server side could not be precisely reconstructed.
Dark Reading also summarizes an Anthropic report, released on Sept. 10 according to the article. Per that summary, Anthropic described the Russian actor Midnight Blizzard reverse engineering a stolen drone vision SDK over “several days” with Claude, and a Chinese campaign that used Claude to reverse-engineer security products in search of bugs. These are Anthropic’s claims as relayed by Dark Reading, and we did not read the original report. Dark Reading also says an open source tool and MCP server called REA became the most popular program on GitHub last month.
Not confirmed: any specific victim, any specific exploited vulnerability, and any CVE. No fixed versions or patches are mentioned.
Am I affected?
The article describes a general trend, not a defect in a product. It is most relevant if you:
- Distribute proprietary client-side software or apps whose logic sits in readable code such as JavaScript. Dark Reading says Slack’s client was easy to unpack this way.
- Analyze malware, where Fishbein says a full reconstruction gives a lot of information.
According to Fishbein, quoted by Dark Reading, “as long as you haven’t [exposed] your server-side infrastructure to the customer, you’re protected.” Whether this holds against determined attackers is not confirmed.
What to do now
- Consider defining which outcomes you want to prevent, for example IP theft or others finding bugs in your code. Dark Reading says only that this “might be” a vendor’s first step before deciding on defensive measures.
- Check what logic you ship to customers versus what stays on your servers, since the researchers could not readily reconstruct Slack’s server side.
- Consider obfuscation, encryption and other anti-analysis measures. Dark Reading notes these have long frustrated defenders, and Fishbein says it is hard to tell how long they will hold as models improve.
- If you do malware analysis, note that the researchers describe an agent-based harness with an extra model for cases where they wanted to avoid executing anything. Treat this as their approach, not a verified recommendation.