acceptodds
Under review as a conference paper at ICLR 2027

HammerEVO: LLM-Driven Evolution of RowHammer Performance-Isolation Defenses

Abstract

Modern DRAM is vulnerable to RowHammer: repeatedly activating a row can flip bits in nearby rows without ever writing them. Safety mitigations stop the flips, but their preventive work—migrations, copies, refreshes—occupies the shared memory-controller service path, and an adversary can trigger that work repeatedly to deny service to benign co-runners. This performance denial-of-service channel is severe yet largely unaddressed: BreakHammer, the existing countermeasure, was evaluated on fixed hammer patterns, not camouflaged ones; real attacks are diverse; and designing a defense demands deep knowledge of each mitigation's mechanism. We present HammerEVO, which applies expert-guided LLM evolution to this problem. The search must cope with a moving adversary, the limited visibility each mitigation exposes, and candidates judgeable only by measurement: domain experts approve hardware extensions and adjudicate measurements, while an LLM proposes attack strategies, probes, and policy code graded by a deterministic, fail-closed simulator campaign. Since evaluating a defense requires real attacks, the campaign evolves both sides. It discovered patterns that degrade benign cores by 13.8–49.1% mean IPC loss (75.6–90.5% worst-core) across multiple mitigation families, which BreakHammer fails to recover, and produced a working performance-isolation policy for each of the eight families: mean benign impact at or below 2.5% (worst-case 7.0%) with near-zero pure-benign overhead, outperforming BreakHammer wherever it engages (worst-core 32–99% reduced to at most 7%) and holding against six published attack families and eight 40-round adaptive red campaigns—for an estimated $50 of API spend and one workstation-week. We release the policies, per-iteration artifacts, and evaluation traces at https://anonymous.4open.science/r/hammerevo-artifact-EB58.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Rowhammer, Hardware safety, Agent, Evolve

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JAyFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HammerEVO, an expert-guided LLM-driven framework for evolving RowHammer performance-isolation defenses. The work addresses the performance degradation caused by conventional RowHammer mitigations, whose preventive operations can be exploited by adversaries to interfere with benign workloads. HammerEVO adopts an attack–defense co-evolution framework, where LLMs generate attack strategies and defense policy candidates, genetic search optimizes attack parameters, and deterministic simulation evaluates their effectiveness under predefined constraints. Domain experts guide the search process and validate design decisions. Experiments across eight RowHammer mitigation families demonstrate that the evolved defenses substantially reduce benign performance degradation compared with existing approaches, including BreakHammer. Additional evaluations examine robustness against published attack families and adaptive adversarial campaigns.

Critical strengths
  1. The paper addresses a meaningful and practical security problem. It highlights the performance denial-of-service risks introduced by RowHammer mitigation mechanisms, extending the focus beyond traditional memory integrity protection.
  2. The attack–defense co-evolution framework is well motivated. The combination of LLM-guided strategy generation, genetic parameter optimization, and simulator-based validation provides a systematic approach to exploring attack patterns and developing corresponding defense policies.
  3. The experimental results demonstrate substantial practical effectiveness. The evaluation covers eight representative RowHammer mitigation families and shows significant reductions in benign performance degradation. The inclusion of published attack families and adaptive adversarial campaigns further strengthens the empirical evaluation.
Critical weaknesses
  1. The independent contribution of the LLM component is not sufficiently established. Although the framework integrates LLM-based strategy and policy generation with evolutionary search, the reported comparison with a simpler search organization does not demonstrate a clear advantage of the specialized HammerEVO framework.
  2. The framework relies considerably on human expertise. Domain experts determine research directions, approve hardware extensions, and make decisions during defense evolution. This makes it difficult to distinguish the contribution of LLM-driven automation from expert-guided policy development.
  3. The generalizability of the evolved defenses remains uncertain. The evaluation primarily relies on cycle-accurate simulation, and the validation of the final policies on genuinely unseen benign workloads and real hardware remains limited.
Questions for authors
  1. The necessity and contribution of the LLM component could be further clarified. The framework combines LLM-generated attack strategies and defense policies with genetic optimization and deterministic evaluation. However, it remains unclear which improvements specifically depend on the LLM rather than conventional search or optimization techniques. The paper also acknowledges that its limited comparison with a simpler search organization does not establish a clear advantage for HammerEVO. Could the authors further clarify the unique role of LLM-based reasoning in discovering effective attack strategies or generating defense policies that would be difficult to obtain through simpler alternatives?
  2. The degree of human involvement and automation should be discussed more explicitly. The framework relies on domain experts to guide the research direction, approve hardware extensions, and select promising defense strategies based on evaluation results. While this design is reasonable for a hardware security problem, it raises questions about the level of automation achieved by the proposed framework. Could the authors clarify which stages require substantial human intervention and discuss how this dependence affects the scalability and reproducibility of the approach across additional mitigation families?
  3. The generalizability and practical deployment implications deserve further discussion. The experimental evaluation demonstrates encouraging results across eight mitigation families under simulation. However, the effectiveness of the evolved policies on real hardware and previously unseen benign workloads remains uncertain. Moreover, the published attack battery is used for development and regression testing rather than as an independent held-out evaluation. It would be helpful for the authors to discuss these limitations more explicitly, particularly the potential challenges of transferring the generated policies to practical memory-controller implementations.
1 comment
AuthorsOpenReview ↗

LLM Contribution, Automation, and Deployment

Thank you for the thoughtful questions. We explain the LLM's role using attack and defense comparisons, describe how expert involvement and automation evolve, and present independent evaluation and implementation evidence below.

Q1. LLM contribution to attack and defense research

The LLM uses mechanism knowledge and feedback to propose attack structures and defense policies; GA searches parameters and simulation evaluates candidates. We select two representative mechanisms: AQUA quarantines aggressor rows, whereas Hydra uses hybrid tracking and preventive refresh.

New experiments: attack generation. Losses are mean/worst-core.

MethodAQUAHydraCost
Fixed-template GABest witnessed: 0.45%/3.17%No witnessed candidate32 / 33 evaluations
Warm-GPBest at 53: 16.85%/76.98%No witnessed candidate96 each
Cold-GPAt 9: 32.12%/79.54%; best at 35: 58.96%/91.98%Best witnessed: 2.78%/6.80%96 each
Direct GA25% mean not reached; best witnessed mean negativeNot tested96 AQUA
Feedback-driven coding agentFirst pass at 16; final 55.16%/91.62%First pass at 6; final 51.67%/89.78%Separate runs; one attacker core
HammerEVOShared attack: 49.2%/79.4%Shared attack: 48.8%/76.6%8 LLM requests + 88 GA evaluations; up to 16 attacker cores

The agent criterion requires a witness and mean ≥25% or worst ≥50%. Conventional searches target historically evolved defenses; HammerEVO's shared attack targets the initial BreakHammer baseline. LLM proposals and GA parameter search support cross-family study of shared structures.

New experiment: tuning an existing defense. We retain BreakHammer's policy logic, add occupancy observation, and sweep 0.5×, 1×, and 2× thresholds. All 18 conditions across six families miss the blue-side targets. Losses are mean/worst-core (%); HammerEVO uses separately frozen attacks.

D0BreakHammer0.5×1×2×HammerEVO
AQUA46.3 / 88.943.0 / 88.643.2 / 88.546.2 / 89.31.40 / 3.01
Hydra20.3 / 78.417.3 / 76.617.2 / 76.617.4 / 76.70.07 / 1.54
RFM17.2 / 76.616.7 / 76.316.7 / 76.416.6 / 76.21.09 / 2.78
PARA17.3 / 76.916.1 / 76.015.7 / 76.216.2 / 76.10.11 / 1.65

These results motivate new evidence, state, and control designs. Experts guide diagnosis and hardware decisions; tools evaluate LLM proposals.

Q2. Human involvement and automation

Automation grows as experts encode experience into HammerEVO:

Early: Experts define representations, interfaces, hardware constraints, and evaluation criteria; the LLM proposes candidates and tools evaluate them.

Middle: Experts analyze feedback, guide defense design, and approve hardware extensions. Observations, calibration rules, and evaluation procedures enter the framework.

Later: Validated interfaces, components, and workflows support new D0 studies and automate established census, calibration, and evaluation steps. Experts assess new mechanisms, signals, and hardware. The iteration figure shows 26 generations for PARA and 2 for Graphene. We will provide policies, attacks, configurations, scripts, and expert-decision records.

Q3. Simulation, independent evaluation, and deployment

Simulation basis. Runtime decision rules, state, and control actions can be implemented and evaluated in simulation, as in BreakHammer and BlockHammer. We use cycle-accurate Ramulator 2.0; its authors checked DDR4 command traces against Micron's Verilog model and reported no timing or state-transition violations in their tests.

New experiments: reserved inputs and hardware. We freeze the eight D1 policies. Literature attacks support development and regression; testing uses 32 new random attacks and 24 reserved SPEC CPU2017 and high-MPKI cloud traces in three mixes. Targets are mean loss ≤5% and worst-core loss ≤10%.

TestPassing casesMaximum mean / worst-core loss
Reserved workloads: 3 mixes × 8 families24/241.34% / 3.14%
Random attacks: 32 traces × 8 families256/2560.97% / 6.00%

No attack-free benign IPC overhead was observed on reserved workloads. We sweep NRH=64–4096, 1/8 channels, LLC=32/64/128 MiB, and 256/512/1024 MSHRs. All 72 LLC/MSHR cases meet both targets; 90/96 NRH/channel cases meet both, all with mean loss below 5% and maximum worst-core loss 15.39%.

New experiment: controller implementation. The added components were implemented in RTL and physically synthesized: cell area is 0.02368 mm² and estimated power is 7.26 mW. Deployment requires integrating the interfaces into a target controller and validating its timing, resources, and workloads.

Reviewer_p3qmOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies attacks that use RowHammer protection mechanisms to slow down other programs. The authors use an LLM to propose attacks and defense policies, test them in a simulator, and revise them with guidance from human experts. The goal is to reduce attack-induced slowdown without harming normal programs. The authors report policies for eight existing RowHammer defenses and show improvements in their main experiments.

Critical strengths
  1. The LLM produces code that is tested in a simulator, giving the authors a concrete way to check whether a proposed defense helps. The evaluation considers both performance under attack and the effect on normal programs when no attack is present.
  2. The experiments cover several defense types and attack patterns. The authors also search for new attacks against the final defenses. This helps determine whether the defenses still work as attacks change.
Critical weaknesses
  1. The method was hard for me to follow. I understand the overall idea: the LLM proposes attacks and defenses, a simulator tests them, and experts guide the next steps. But I couldn't follow what happens in a single iteration. An example showing the expert’s instructions, the LLM’s output, the test results, and the next change would help. It would also make it easier to see what the LLM contributed and what the experts contributed.
  2. I am not sure how to interpret the results on adaptive attacks. The abstract says the defenses hold against adaptive attacks, but Section 4.5 reports: "the best measured mean loss is 19.0%" and "the best AQUA and Hydra attacks still inflict 55.4%/55.2% worst-core loss". The authors explain that these attacks do not meet their definition of a successful attack, but they still cause substantial slowdown. I would like a clearer explanation of what “holding” means here. The abstract and conclusion should also clearly explain the remaining slowdown.
Questions for authors

Could you walk through one full iteration of the method, showing the expert’s instructions, the LLM’s output, the simulator’s feedback, and what changes next?

1 comment
AuthorsOpenReview ↗

Adaptive Attacks and Red–Blue Iteration

Thank you for highlighting the residual slowdown and requesting a concrete walkthrough. We clarify the adaptive-attack results, report the strengthened defenses and subsequent evaluations, and illustrate a complete red–blue iteration below.

Weakness: adaptive attacks and residual slowdown

In the submission, “holding” meant that adaptive search found no attack meeting the red-side criterion: a mechanism witness and at least 25% mean loss. Nevertheless, substantial slowdown remained: AQUA suffered 19.02%/55.44% mean/worst-core loss; a Hydra candidate suffered 13.91%/55.17% and missed the witness gate. Both exceed the blue-side targets of mean loss ≤5% and worst-core loss ≤10%.

We treated the defense as passing too early when red-side success was not reached. Even below the red threshold, observed slowdown reveals weaknesses that can guide further defense design. Red–blue iteration uses these findings to revise defenses and challenge them again.

New experiment: hardened defenses. We evaluate earlier and hardened policies under the same attacks and configurations. The table reports mean/worst-core loss (%):

D0EarlierHardened
AQUA19.34 / 55.481.40 / 3.01
Hydra12.04 / 44.230.07 / 1.54

New experiment: subsequent adaptive search. Starting from existing attacks, we evaluate 32 attack candidates each for AQUA and Hydra. The defenses meet both blue-side targets against all measured candidates. Maximum mean and worst-core losses are taken separately and may occur in different cases:

D0PassingMaximum mean / worst-core loss
AQUA32/321.40% / 3.89%
Hydra32/320.49% / 3.12%

Q1. A complete red–blue iteration

The following Hydra cases summarize the expert tasks, LLM proposals, simulator feedback, and subsequent decisions from research records.

Blue side: expert-guided observation and policy design.

ActorAction and feedback
ExpertRequest a policy that identifies attack interference and protects benign programs.
LLMPropose RCT-LATCH, using Hydra row-counter pressure to control requests.
SimulatorThe proposal misses the 5%/10% targets; a census shows weak attack–benign separation. Further proposals using existing hardware also fail.
ExpertRequest analysis of policy requirements, signals, and hardware.
LLMPropose controller-occupancy observation.
ExpertRequest hardware cost and compatibility analysis, including whether DRAM changes are needed.
LLMIdentify a CPU-controller-side implementation.
ExpertApprove the addition and request simulator changes and validation.
LLMModify and validate the simulator, build OCC-LATCH, and evaluate it.
SimulatorThe eight-channel smoke test gives 0.97%/2.75% mean/worst-core loss, with zero attack-free overhead.

Red side: LLM proposals and GA attack search.

ActorAction and feedback
ExpertRequire D0 preventive actions and specify seven operators and parameter ranges.
LLMGenerate eight strategies, each with a mechanism description, 13 parameters, and nine operators. MVP combines probing, hammering, bursts, rotation, bank expansion, mean poisoning, and dummy accesses.
FrameworkRun six probes per proposal; 5/8 templates pass screening.
GA and simulatorEvaluate 88 parameter instances by compiling, simulating, and measuring benign-core IPC.
FrameworkCompile the selected nine-phase attack for 16 cores with 126,473 records per core; test eight D0s with D0-only and D0+BreakHammer deployments, then freeze it.
ResultAll 8/8 D0s meet the criterion of mean loss ≥25% or worst-core loss ≥50%. Hydra reaches 48.8%/76.6% in both deployments. The run uses 8 LLM calls, 88 GA evaluations, and 134 simulations.
Reviewer_tswBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an expert-guided LLM framework for automatically discovering attacks and designing defenses against RowHammer attacks. Specifically, the paper introduces red-side evolution where an LLM proposes attack strategies, genetic algorithms optimize their numerical parameters, and a deterministic simulator evaluates their effectiveness Then the blue-side evolution uses An LLM generates and iteratively refines defense policies based on simulator feedback, while human experts guide research decisions and approve hardware extensions. Experimental results show the effectiveness of identifying new attack vulunerability.

Critical strengths

+) Effective application of LLM-driven attack–defense co-evolution. The framework combines LLM-generated attack strategies, genetic parameter optimization, LLM-generated defense policies, and deterministic simulator feedback.

+) Strong empirical performance across diverse mitigation mechanisms.

Critical weaknesses

-) Methodological novelty is weak. The proposed framework primarily combines established techniques, e.g., LLM-based program generation, evolutionary optimization, and execution-feedback refinement. Red/blue teaming is also not a new concept. It seems an LLM-based zero-day attack discovery on a specific type attack.

-) Limited generalization and robustness evidence. Most evaluations rely on simulator configurations and attack workloads used during development.

Questions for authors

a) Can the authors provide controlled comparisons against simpler approaches, such as standard LLM coding agents with iterative feedback?

b) How well do the discovered policies generalize to unseen attack patterns, benign workloads, and hardware configurations?

c) How well does the proposed method be applied to other attacks?

1 comment
AuthorsOpenReview ↗

Baseline Comparisons and Generalization

Thank you for the questions. We address baseline comparisons, generalization, and applicability below. All losses are benign IPC losses.

Q1. Why LLMs, and how does HammerEVO compare with coding agents?

LLM role. Attack research combines access structures with numerical parameters; defense research selects observations, state, and controls while protecting benign performance. LLMs use mechanism knowledge and feedback to propose attack structures and executable policies. Experts guide research and hardware choices, GA searches parameters, and simulation evaluates candidates. On eight frozen challenges, the evolved defenses reduce mean loss from 38.5%–51.1% with BreakHammer to at most 2.5%, with near-zero attack-free overhead.

New experiment: attack generation versus a feedback-driven coding agent. Both attack D0 + BreakHammer. Independent coding-agent runs use up to 50 evaluations per D0: 7/8 meet the attack criterion, averaging 26.25 evaluations across runs. HammerEVO finds a shared attack covering 8/8 D0s with 8 LLM requests and 88 GA evaluations, averaging 48.1% mean loss across families. The agent uses one attacker core; HammerEVO allows up to 16. The agent criterion requires a mechanism witness and mean loss ≥25% or worst-core loss ≥50%.

New experiments: defense generation versus coding agents and autoresearch. Autoresearch methods receive a frozen attack; coding agents use implemented observation/action interfaces.

MethodEvaluationsResult
OpenEvolve62Runnable; best mean 54.2%
ShinkaEvolve~50Runnable; best mean 54.2%
Self-Debug—No compilable candidate
AIDE / AI-Scientist-v2—Crashed
Coding agents≤32 per D07/8 pass; successful runs average 15.14 evaluations
HammerEVOFrom 26 PARA to 2 Graphene generationsEight defenses reach ≤2.5% mean loss

HammerEVO retains censuses, observations, and policy components for reuse. Cross-family experience also supports a shared D1 framework.

Q2. Generalization to unseen inputs and hardware

We freeze the policies; targets are mean ≤5% and worst-core ≤10%. Maxima are independent and may arise in different cases.

New attacks and reserved workloads. All 256 cases from 32 random traces × 8 D0s pass, with maximum mean/worst loss 0.97%/6.00%. We also reserve 24 SPEC CPU2017 and high-MPKI cloud traces in three mixes: all 24 D0–mix cases pass, with maxima 1.34%/3.14% and no measured attack-free overhead.

D0Random: max mean/worst (%)Reserved: max mean/worst (%)
AQUA0.87/4.831.26/3.03
Graphene0.62/4.480.39/1.30
Hydra0.97/4.980.35/1.29
PARA0.89/6.001.30/3.14
PRAC0.67/2.791.34/2.58
ReGA0.88/5.401.24/2.81
RFM0.75/3.331.28/2.88
TWiCe0.88/5.401.24/2.81

Hardware configurations. We sweep NRH=64–4096 with 1/8 channels and LLC=32/64/128 MiB with 256/512/1024 MSHRs. All 72 LLC/MSHR cases pass. Of 96 NRH/channel cases, 90 pass both targets; all have mean <5%, and maximum worst loss is 15.39%.

D0NRH/channels: pass; max mean/worst (%)LLC/MSHR: pass; max mean/worst (%)
AQUA12/12; 1.96/9.679/9; 1.17/2.22
Graphene12/12; 4.08/9.969/9; 0.68/2.95
Hydra11/12; 3.52/13.879/9; 0.36/3.24
PARA10/12; 3.56/15.399/9; 1.11/5.98
PRAC12/12; 1.88/4.609/9; 1.19/4.60
ReGA10/12; 4.49/12.339/9; 1.42/7.69
RFM11/12; 2.72/12.119/9; 1.43/4.67
TWiCe12/12; 2.37/6.649/9; 1.42/7.69

Q3. Applicability to other attacks

For a new problem, experts define attacker operations, representation, evaluation environment, defense observations, and legal actions; LLM proposals, search, and feedback support attack and policy research.

New experiment: random memory-contention attacks. We test 32 traces per D0 with physical preventive actions on and off; tracking, scheduling, and counters remain active. Values are family-mean / maximum worst-core loss (%).

D0D0 offD0 on
AQUA0.52/4.830.52/4.83
Graphene0.47/5.400.47/5.40
Hydra0.47/4.980.47/4.98
PARA0.47/5.400.48/6.00
PRAC0.39/2.790.39/2.79
ReGA0.47/5.400.47/5.40
RFM0.47/5.400.37/3.33
TWiCe0.47/5.400.47/5.40

All 256 D0-off cases pass both targets. Across families, mean loss falls from 33.4%–34.5% without D1 to 0.39%–0.52% with D1, demonstrating isolation under the tested contention attacks.

Comments on the paper

1 comment

Dear Reviewers,

Please take a moment to read the authors’ rebuttal and participate in the discussion.

Thanks,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.