Understanding AI Beyond the Headlines (4): Did AI Discover It? A Closer Look at PHP2Shell and Node2Shell

Part 4 of 4 — Understanding the technology behind the headlines.

AI ExplainedOne of the underlying problems behind PHP2Shell was already known in 2024. In 2026, researchers using AI demonstrated how it could be exploited to run commands on affected systems.

That history changes how we should read a claim that “AI discovered the vulnerability.” The later work mattered, but the problem did not begin with it.

Node2Shell raises a related question: what counts as a discovery when faulty code has already been corrected, but nobody recognised its security implications?

PHP2Shell: what was already known

According to the PHP Foundation, early evidence of the out-of-memory problem was found in 2024 during work following an external audit. Researchers had considered possible exploitation, but concrete attack paths were missing. Fixing the issue involved difficult tradeoffs, and progress stalled. [1]

In August 2026, Calif reported remote code execution through the same root cause. In plain English: researchers had demonstrated how the memory-handling problem could let an attacker run code. Further reports and attack paths followed. [1]

The distinction is straightforward. The underlying problem was known; practical exploitation had now been established.

That is valuable research. It can turn a suspected risk into something developers can reproduce and address. It does not make the original finding a new discovery.

What did AI contribute?

Calif describes AI-assisted source-code analysis that identified vulnerable behavior and traced ways to exploit it. The work also included additional findings. [2]

So “AI merely wrote the exploit” would oversimplify the research too. Investigating a codebase, finding a reachable attack path, and constructing a demonstration are substantial contributions — even when an earlier investigation had identified the underlying problem.

PHP2Shell also covers two related vulnerability classes. The documented 2024 history concerns the out-of-memory issue; it should not automatically be extended to every finding under the PHP2Shell name. [3]

The accurate account needs both the earlier discovery and the later work.

Node2Shell: corrected code, unrecognised danger

Node2Shell concerns Brotli, a decompression library used by Node.js and other software. Node.js credits Khai Tran and Sylvie Mayer of OpenAI Codex Security with reporting the vulnerability. [4]

According to OpenAI’s public account, the root cause had already been corrected in Brotli’s source code in February 2026. Its security significance had not been recognised, and a newer release was not yet available at the time described. [5]

This differs from the PHP history. A code correction does not establish that a security vulnerability was already understood and disclosed. Recognising that a bug enables an attack can itself be a discovery.

We therefore cannot dismiss everything after the earlier fix as simply writing an exploit. We need to identify what the later investigation established.

Tenable’s test: the result worked, the explanation didn’t

During PHP2Shell testing, Tenable encountered an AI agent that achieved command execution and reported success against PHP2Shell. A researcher checked the steps and found that it had actually exploited a different weakness deliberately present in the lab. [6]

The agent had produced a real result and attributed it to the wrong cause.

That makes verification part of the research, not a formality after it. A successful demonstration must show that the claimed vulnerability caused the result.

Credit the actual achievement

These cases show why “AI discovered it” needs more detail. An existing problem can gain a new exploit. A corrected bug can reveal a previously unrecognised security risk. A convincing demonstration can prove something other than what its author claims.

My conclusion is simple: give AI-assisted research credit for what it establishes, and preserve the history that came before it.

What was already known? What is new? And who checked that the evidence supports the claim?

Sources

Based on reports reviewed through November 1, 2026.