security auditing teams

Abstract: Automated jailbreak discovery aims to find prompts that elicit prohibited behaviors from safety-trained language models, helping developers identify vulnerabilities before deployment. One class of approaches uses reinforcement learning to train an investigator model to generate prompts that elicit a specified behavior from a target model. However, training directly against a safety-trained target is difficult because nearly all prompts are refused, leaving the investigator with a sparse reward signal. Prior work addresses this with the propensity bound (PRBO), a proxy reward estimated by sampling responses from a more compliant target, and then adjusting for how likely those responses would have been under the original target. We introduce Curriculum of Models (CoM), an alternative method that trains the investigator against a sequence of modified versions of the target, progressing gradually from a version that readily exhibits the specified behavior to the original target. For jailbreak discovery, training begins against a target version that rarely refuses harmful requests, and advances to the next, harder version once the investigator reliably elicits the behavior from the current one. We evaluate CoM on elicitation tasks involving CBRN materials, explosives, and illicit drugs, using two open-weight targets, gemma-4-31b-it and gpt-oss-120b. We compare against PRBO and direct training on the original target without a curriculum. When the investigator is trained to elicit a single behavior, CoM finds successful jailbreaks more reliably than either baseline. When trained on multiple behaviors simultaneously, the CoM-trained investigator can elicit held-out behaviors, though PRBO-trained investigators achieve higher success rates. Our results suggest that a simple curriculum over target models offers a promising approach for addressing sparse rewards in rare-behavior elicitation.

Abstract: Multi-agent systems built on large language models (LLMs) are increasingly deployed in high-stakes settings such as finance, healthcare, and software engineering, where agents coordinate through natural-language messages. The same channels, however, let colluding agents exfiltrate confidential information or coordinate unauthorized actions, and steganography can hide such communication inside outputs that look ordinary to an auditor reading the transcript. Existing provably undetectable LLM steganography protocols are not suited to realistic deployments. High-capacity schemes assume a symmetric setting where the receiver can reproduce the sender's output distribution, the state-of-the-art protocol for asymmetric agents has very low capacity, and most approaches rely on a pre-shared secret key. We make the threat of undetectable agent collusion concrete with Codetta, a high-capacity steganographic protocol for independently deployed agents in realistic asymmetric settings. Codetta combines a shared public model that estimates the communication channel, a sampling mechanism that preserves the sender's output distribution, and an adaptive error-correcting code. It further removes the pre-shared key through a steganographic key exchange that lets independently deployed agents establish a shared key while keeping the transcript computationally indistinguishable from ordinary model outputs. Across three agent workloads and three sender models, Codetta achieves up to $94\times$ the capacity of the state-of-the-art asymmetric protocol, and its key exchange establishes a shared key with about 80k visible tokens at an empirically certified failure probability of at most $4.1\times 10^{-3}$. These results show that effectively undetectable collusion is becoming feasible between independently deployed agents, so auditing must go beyond inspecting communication transcripts.

Abstract: Code generated by large language models passes security checks at a rate that has barely moved in four years. In regulated backends, the defect classes that matter most are money arithmetic, time handling, retry safety, and access control. Teams answer with instruction files, yet the largest controlled study of instruction files to date found no benefit. This paper tests a narrower idea: generated code improves when the prompt carries a specification, a fixed preamble stating what must be true of the result. We pre-registered hypotheses, refuters, analysis code, and a one-shot generation rule, then ran 50 realistic backend tasks from finance, healthcare, and insurance practice through five frontier models from five vendor lineages, each task twice: bare, and preceded by a 267-word filled specification frame. Nine deterministic AST-based checkers scored the outputs. The Bandit security scanner, which knows nothing of the frame, scored them independently. The frame reduced defects in all five models (mean reduction 0.16 to 0.70 findings per task, every Holm-adjusted sign test significant, every bootstrap confidence interval excluding zero). Where the arms differed, the frame arm won 95 of 100 times. It never made any model worse in any domain. Bandit found 53 medium-or-high issues in the bare arm and 11 in the frame arm, in the same direction for every model. The effect was largest where a model's unprompted defaults were weakest: the frame supplies the discipline a model lacks. All 500 outputs, prompts, checkers, scoring code, and the pre-registration are published with a DOI, so any team can re-derive the result without trusting the author.

Abstract: Proofs of UNSAT are a standard primitive in formal verification and software assurance. In many real-world settings, the proof itself encodes proprietary or security-sensitive information, making public disclosure undesirable. Zero-knowledge certification of UNSAT addresses this tension: it enables a prover to convince a verifier that no satisfying assignment exists, without revealing anything about the underlying proof beyond its validity. Luo et al. recently introduced ZkUnsat, a protocol that achieves this goal by proving the validity of a weakened resolution proof in zero knowledge. ZkUnsat demonstrates the feasibility of zero-knowledge certification; however, its scalability to larger, real-world instances is constrained by substantial prover memory overhead, limiting its real-world applicability. Motivated by advances in UNSAT proof formats such as LRAT, which enable efficient plain-text verification, we present a preprocessing technique that improves the efficiency of ZkUnsat without introducing additional leakage. Our approach normalizes the proof so that each derived clause is justified by a resolution chain of fixed public length k. This eliminates chain-length leakage and reduces prover memory usage. With k = 16, our method certifies roughly 62% more instances than baseline ZkUnsat on the SAT 2002 competition benchmarks. Furthermore, for an equivalent number of certified instances, the memory footprint drops to under 25% of that required by the baseline.

Abstract: The Open Network (TON), with a peak market capitalization exceeding $20 billion and over 175 million activated on-chain addresses, relies on the TVM (TON Virtual Machine) to execute smart contracts. TVM uses first-class continuations with savelists to manage control flow and register state across continuation invocations. Since savelist-captured registers allow data to flow across continuation boundaries without passing through the operand stack, bytecode-level analyses cannot construct complete data flow tracking without explicitly modeling savelist semantics. We present TasmScan, the first bytecode-level static analysis framework for TVM that enables cross-continuation data flow reasoning without requiring source code. TasmScan models savelist semantics via forward register analysis with a formal over-approximation guarantee for exact-resolved save sites and locally tracked register definitions, then lifts bytecode into TASIR, a typed intermediate representation, and performs path-sensitive taint analysis with context-aware sources to detect defects. We evaluate TasmScan on 2,921 contracts from the TON verifier registry and a labeled benchmark of 208 contracts with human-confirmed ground truth. On the full corpus, TasmScan resolves 294,546 dynamic continuation targets with 100% precision; ablation confirms that savelist propagation is essential for resolving indirect register calls that depend on cross-continuation register passing. On the benchmark, TasmScan detects 95.3% of defects across five classes with 96.8% precision. A 366-pair stratified sample from the full corpus estimates 85.8% overall precision. TasmScan offers a 17x median speedup over the state-of-the-art symbolic-execution baseline, and in the path-analysis comparison completes 100% of analyses with zero crashes or timeouts.

Abstract: Software agents with Large Language Models (LLMs) are designed for Automated Program Repair (APR) tasks, raising the possibility that, in the near future, APR agents will fix bugs automatically without much human intervention. Can we trust an APR agent to produce both functionally correct and secure code in such situations? What if attackers target production APR agents with adversarial issues that seem benign but may influence the agents to produce correct but insecure code? In this paper, we took a first step towards answering these questions by conducting an empirical study. First, we created SWEADV, a benchmark of 750 adversarial issue descriptions constructed from 150 repair tasks in SWE-bench Verified. For each repair task, we created five adversarial issue descriptions, one for each attack type: command execution, deserialization, path traversal, denial of service, and weak hashing. Second, we evaluated mini_swe APR agents from three LLM backends on SWEADV: GPT-5-Mini, MiniMax-M2.5, and DeepSeek-R. We found that on average, adversarial issue descriptions can induce malicious behaviors with successful repair in 51.7% of cases. Third, we investigated whether typical detection mechanisms are sufficient to prevent such malicious patches from being accepted. Pre-repair detection with LLM-as-judge on the adversarial issue descriptions resulted in an average detection accuracy of only 62.3%. Post-repair detection on adversarial APR patches using static analysis tools and LLM-as-judge achieved average detection accuracies of only 39.4% and 55.4%, respectively. We conclude that autonomous APR agents cannot be trusted yet in production deployment, given their susceptibility to adversarial attacks.