When processor implementation details are unavailable, black-box microarchitectural reverse engineering helps security researchers identify hardware conditions related to security vulnerabilities. However, this process requires substantial manual effort to design experiments, develop test programs, and analyze results for each processor. We present ProcReverser, a multi-agent system that automates this process. ProcReverser uses execution feedback to improve experimental plans and test programs and identifies microarchitectural properties from measurements. We develop a comprehensive suite of property identification tasks focused on hardware mechanisms relevant to Spectre-class attacks, covering speculative execution, branch prediction, memory dependence, and cache behavior. We evaluate ProcReverser on the XuanTie C910, the gem5 RISC-V O3 CPU model, and the SiFive P550. ProcReverser achieves Pass@1 and Pass@5 scores of 81.3% and 93.8%, respectively, on C910, and 83.0% and 100.0% on gem5 O3. On the SiFive P550, ProcReverser further characterizes microarchitectural properties covered by the proposed task suite. These findings highlight the potential of automated microarchitectural reverse engineering to support processor vulnerability analysis and defense evaluation.
Figures & tables
Fig. 1: ProcReverser’s experimental workflow. Solid arrows show the main workflow. Dashed arrows show feedback paths for error reporting, repair, and revision.
Fig. 2: Example input for the C5 task.
ID Property task P1 Cache-signal suppression by tested instructions and transient-window type P2 Store-loop cycle overhead of tested barriers P3 Barrier effectiveness at tested instruction spacings P4 Cache-signal suppression from non-counter CSR reads R1 RAS capacity in return addresses R2 RAS overwrite on overflow and fallback when empty R3 Privilege-related predictor tagging and partitioning R4 Direction-predictor state persistence across address spaces or process switches R5 Direction-predictor state decay under time or branch interference R6 Branch-history dependence of direction prediction R7 BTB capacity by level or target configuration R8 Branch-history parameters and observable dependence R9 Predictor indexing, aliasing, and cross-context interference
ID Property task R10 BTB entry insertion and update behavior W1 Transient-execution window length under tested triggers W2 Effects of tested instruction layouts on transient windows C1 Availability of user-mode timing and counter interfaces C2 Available observation channels and cache-maintenance operations C3 Cache capacity and organization of characterized levels C4 SMT support per physical core C5 Presence of a memory-dependence predictor C6 Hardware data prefetching enablement and observed behavior C7 Cross-core cache sharing (L2/LLC) F1 Enablement of tested branch-prediction components S1 Post-fault transient cache traces S2 Software control and observed clearing of predictor state
Load/store fault tests. Read-ordering fences, fence.i , and counter CSR reads suppress the cache signal. fence w,w , fence.tso , and lr.d do not. [C]
R1
RAS capacity: 12 entries . [A,B,C]
R2
Overflow: circular overwrite. Underflow tests: no BTB fallback or straight-line speculation. [B,C]
R3
RAS/Ind-BTB: privilege tags. BHT: no explicit ASID/privilege index key. [B]
R4
PHT training crosses address spaces . [C]
R6
Branch history contributes to direction indexing. [B]
TABLE II: C910 reference answers.
ID
Reference configuration and behavior
P4
fcsr , fflags , and frm reads suppress the cache signal from wrong-path loads in the tested SE branch window.
R1
RAS: 8 entries in the baseline, 16 in the alternate instance.
R2
Overflow: overwrites the oldest entry. Underflow: stale address or empty slot. If empty, keep the BTB target if present, else fall through.
R3
LocalBP has no partitioning by privilege level .
R4
LocalBP state persists across same-core process switches .
R5
LocalBP counters do not decay with time alone .
TABLE III: Reference answers for 20 gem5 O3 tasks.
Fig. 3: Identification outcomes for the 16 C910 reference tasks. A1–A5 denote five independent attempts; each cell represents one scored task attempt.
Fig. 4: Identification outcomes for the 20 gem5 O3 reference tasks. A1–A5 denote five independent attempts; each cell represents one scored task attempt.
Configuration
C910
O3
Mean cost
Pass@1
Pass@5
Pass@1
Pass@5
Plans
Exec.
Tokens (M)
Single agent
17.5%
37.5%
18.0%
40.0%
7.06
75.09
3.34
Single agent + DR
22.5%
50.0%
25.0%
45.0%
7.83
47.66
2.87
Five roles
68.8%
87.5%
72.0%
90.0%
4.44
40.36
13.11
Five roles + DR
81.3%
93.8%
83.0%
100.0%
5.66
22.55
12.88
TABLE IV: Ablation of role specialization and DR.
Fig. 5: (a) RAS depth identification on gem5 O3. (b) Hardware prefetching characterization on P550. Error bars indicate 95% confidence intervals for the mean hit rates.
ID
Measured conclusion
P1
All nine user-mode barriers listed in Table I showed attenuation of the measured cache signal in both memory-dependence and fault-triggered windows.
P2
In a store loop, each execution of fence r,w , fence w,w , and fence.i added about 3, 10, and 46 cycles, respectively. fence rw,rw , fence r,r , fence w,r , and fence.tso each added about 13 cycles.
P3
fence.i suppresses the fault-window cache channel at tested spacings of 0–16 instructions.
P4
fcsr reads do not suppress the tested fault window.
R1
16-entry RAS.
R4
Same-core direction-prediction state persists across processes.
TABLE V: Property characterization results on SiFive P550.
Modern processor verification struggles to reach deep architectural states due to the inefficiencies of traditional mutation-based fuzzing. We propose HiFuzz, a novel hierarchical reinforcement learning framework that replaces mutation with a structured, two-layer generation process: a Program Agent for global layout and a Basic Block Agent for precise instruction filling. To overcome reward sparsity, HiFuzz integrates an adaptive coverage reward mechanism and a semantic-aware basic block encoder providing intrinsic feedback. Extensive evaluations on three real-world RISC-V cores demonstrate that HiFuzz significantly outperforms state-of-the-art fuzzers in coverage and bug detection.
Ya Wang, Hanwei Fan, Zhenguo Liu +4
Hong Kong University of Science and Technology (HKUST), Hong Kong · The Hong Kong University of Science and Technology (Guangzhou), China
As software systems grow increasingly complex, automated vulnerability repair (AVR) remains difficult because the materials available to a repair system are usually failure artifacts rather than repair guidance. Traditional analysis techniques can provide suspicious locations, reduced triggers, or constraints, but they are costly to configure across repositories and seldom directly actionable for patch generation. Recent LLM-based agents can edit and validate repository-level patches, and experience-based systems can reuse prior repair traces or demonstrations, but they still need current-instance evidence that turns a broad, symptom-level failure report into a concrete repair decision. We present ContraFix, an agentic AVR framework that constructs such evidence through contrastive runtime analysis. Starting from a failing witness, ContraFix generates nearby failing and non-failing variants, executes them through aligned probe sites, and compares their runtime states to infer the repair boundary and guide source-level patching. Each candidate patch is accepted only after build and validation. ContraFix also stores validated repair episodes in a dual-track skill base, reusing mutation skills to construct useful variants and correction skills to refine failed patches. On SEC-Bench, ContraFix with GPT-5-mini achieves resolution rate of 92.0% over three repeated runs and an average resolution rate of 91.8% +/- 0.8. On PatchEval, it resolves 73.8% of 225 Go, Python, and JavaScript instances. A semantic audit of benchmark-validated SEC-Bench patches shows that 58.2% of ContraFix's patches are semantically correct, compared with 31.3% for the strongest baseline, indicating that the proposed framework improves semantic correctness beyond benchmark validation.
Simiao Liu, Fang Liu, Peiding Wang +4
Beihang University Beijing, China · The University of Hong Kong Hong Kong, China
AI agents now report vulnerabilities faster than maintainers can review them. Reports often depend on security properties specific to the application, and require considerable human labor to process. To mitigate this, we introduce a framework for evaluating vulnerability reports via probes, executable checks of security properties. A reported exploit is evaluated by replaying it against the application and running the probes: a triggered probe indicates both that the exploit succeeded and which security property it violated. As a probe encodes a security property rather than a known vulnerability, it can detect vulnerabilities that were not known when the probe was written. We instantiate the framework as MobileCybench, a benchmark for vulnerability discovery by AI agents in 13 Android applications, with 495 probes written and reviewed by the authors. We evaluate 5 coding agents (OpenCode with GPT-5.5, GPT-5.6-Sol, and GLM-5.2; Claude Code with Opus 4.8 and Opus 5) under 4 settings: as a malicious app on the victim's device or as a remote attacker with a low-privilege account, each with either only an obfuscated APK or access to the application's source code. Given only the obfuscated APK, the top agent, OpenCode with GPT-5.6-Sol, triggers probes in 53.8% of applications in the malicious-app setting and 16.7% in the remote-attacker setting. With source code, the trigger rate across all agents and both attack settings increases from 28.8% to 32.8%. Building and running the benchmark surfaced 23 previously unreported vulnerabilities, the majority of which have been confirmed by maintainers.