Training Language Models to Discover RowHammer Vulnerabilities in DRAM

Raising $5,000active

Build an open-source RL environment using Ramulator and real disturbance data to post-train language models to exploit simulated DRAM RowHammer vulnerabilities, plus write-up/blog and trained models.

Project Details

We will build a reinforcement learning environment in which language models can be post-trained to exploit RowHammer vulnerabilities in simulated DRAM. In RowHammer attacks, an adversary rapidly accesses (“hammers”) DRAM rows, which can cause bits to flip in physically adjacent memory. RowHammer has been demonstrated to enable privilege escalation over the network and extract keys from a secure enclave.

We have already built a proof of concept of this environment, the code for which is available here. Our environment uses Ramulator 2.1 (https://arxiv.org/abs/2606.13844) to simulate DRAM, and we fit a read-disturbance model using real data, which allows us to simulate RowHammer-vulnerable DRAM with a high degree of realism.

The concrete outputs of our work will be:

An open source environment for training models to exploit RowHammer vulnerabilities in simulated DRAM
A write-up of our environment and a blog post arguing that humanity is dropping the ball on automating hardware security research
Models post-trained in our environment

We also plan to build other hardware security environments, though this may end up out of scope for this grant depending on how difficult building the RowHammer environment ends up being.

Theory of Impact

The frontier labs are investing heavily in securing software (e.g., Glasswing and Daybreak), but hardware security has received little attention. As a result, hardware may end up significantly more vulnerable than software. In the near future, an attacker could plausibly post-train an open source model to exploit particular hardware vulnerabilities and use this model for malicious purposes. Our project may seem counterintuitive because we are providing tools for making models better at exploiting hardware vulnerabilities, but we believe that building these environments now can alleviate the current overhang in the ability of language models to exploit hardware vulnerabilities. These environments are not prohibitively difficult to build and there is a large incentive to exploit hardware vulnerabilities, so it seems unlikely that malicious actors will not do so in the future. Building these environments now allows them to be used defensively.

People

Jaray Foo

Team Member

Luke Marks

Team Member

Funding Details

Start Date: -
End Date: -
Expected Duration: 1 month
Funding Raised to Date: -
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: -
Fiscal Sponsor: -

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$5,000

Ideal

$10,000

How the money will be spent

The money will almost entirely be spent on compute for training large open weight models in our environment. We may spend a small amount of the money on API credits.

The ideal amount will enable us to train models to exploit simulated DRAM despite RowHammer mitigations (e.g. TRR, ECC, RFM), and will allow us to support additional DRAM standards (HBM3/4, GDDR7) which could require gathering read-disturbance measurements ourselves. The minimum amount will enable us to train models on a small number of DRAM standards, and likely without RowHammer mitigations.

Discussion

No comments yet. Be the first to share your thoughts.