OpenAI claims to have solved reverse engineering

[chants]

We also tested Astra on SRE-Bench15, a benchmark that measures whether models can reverse engineer software binaries to understand its core logic without access to raw source code. Astra solved 88.0% of tasks in a single attempt and 99.2% within four attempts, compared with 55.9% and 68.7% for GPT‑5.6 Sol, respectively.

https://openai.com/index/gpt-6-astra/

[Kurapica]

I remember it was supposed to cure cancer too ?

[th3tuga]

Originally Posted by Kurapica

I remember it was supposed to cure cancer too ?

It is doing that too since past year!

[chants]

This paper is worth a mention in regards to SRE benchmarking

The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark
https://arxiv.org/abs/2608.11469v1

[WhoCares]

Can it solve VMP ?

It can learn the leaked source code of VMP.

[dyers eve]

Originally Posted by WhoCares

Can it solve VMP ?

It can learn the leaked source code of VMP.

Both Claude and OpenAI can completely devirt the VMP. Tried this last year. It will use up lots and lots of tokens.
So it is not a good idea to try it from personal accounts.

[Kurapica]

Originally Posted by dyers eve

Both Claude and OpenAI can completely devirt the VMP. Tried this last year. It will use up lots and lots of tokens.
So it is not a good idea to try it from personal accounts.

Any real life examples? reports ? analysis ?

[Fyyre]

Would make more sense to have it build a devirt than devirt a single application... also, I'm amazed this kind of request (especially if to the chatbot) did not trigger a refusal.

[th3tuga]

Originally Posted by Fyyre

Would make more sense to have it build a devirt than devirt a single application... also, I'm amazed this kind of request (especially if to the chatbot) did not trigger a refusal.

I also was able to get it done last year. I'm not sure how the other guy did it, but this is how I managed it.

To rebuild the final solution, I manually fed the decompiled code from IDA into the chatbot in logical chunks. This was a grueling process that took several hours over a few days. Because of the massive effort and time required to input the code piece by piece, I have been searching for a more automated solution.

I also noticed that OpenAI's models were much less likely to refuse requests last year, before the introduction of Codex. While I haven't tried recently, I suspect that bypassing these refusals would be significantly more difficult now.

Refusals are significantly less when requested from Enterprise OpenAI accounts. Last year I had access to a leaked one after paying a trivial amount in BTC. I find it significantly more difficult to come across such usuable leaks these days.