[ { "slug": "crypto-zk", "name": "Cryptography and ZK specialists", "short": "cryptography and ZK scanning", "title": "AI bug-finding tools for cryptographic code and ZK circuits: zkao, zk-skills, zkCraft and others", "summary": "Tools built for the code that general scanners handle worst: finite-field arithmetic, constraint systems, pairing libraries, MPC and post-quantum implementations. zkao (zkSecurity) is the only product in this class with public, upstream-confirmed critical findings; zk-skills makes the same audit patterns available as open-source agent skills; zkCraft adds LLM guidance to circuit fuzzing.", "intro": [ "Cryptographic bugs rarely look like the memory-safety and injection patterns that security models are trained on. A missing subfield membership check, a rogue-key attack on aggregate signatures, or an unconstrained signal in a circuit is only a bug relative to a mathematical specification the tool must understand. The tools here encode that knowledge, either as expert-maintained skills, as multi-agent workflows that build a model of a module's invariants before hunting for violations, or as oracles for fuzzers.", "The public evidence base is small but concrete: in 2026 zkao found seven bugs in Cloudflare's CIRCL, a critical soundness bug in OpenVM's pairing library and four zero-days in Bron Labs' library, all confirmed and fixed. zkSecurity's write-ups also document the failure modes: severities mis-rated in both directions, triage of more than a thousand candidates as the bottleneck, and plain frontier-model runs missing what the harnessed workflow found." ], "choose": [ "Circom, Leo, Rust or Go cryptographic code with a budget for continuous scanning: **zkao**.", "A first pass you can run yourself with Claude Code, Codex or Cursor on Circom: **zk-skills / circom-auditor**.", "Automatic, specification-free circuit checks alongside AI: **zkFuzz** with **zkCraft** guidance (and Picus or CIVER, covered on the formal verification side).", "Solidity plus Circom, Halo2 or Noir in one product: **Savant Chat** lists all of them; check its ZK results before relying on it.", "Whatever tool you run, budget human cryptographer time for validation; every vendor in this category says so." ], "page": "https://agentsast.com/categories/crypto-zk/", "tools": [ "zkao", "zk-skills", "zkcraft", "ai-grinding-cryptanalysis" ] }, { "slug": "frontier", "name": "Frontier-lab and general scanners", "short": "general-code scanning", "title": "Frontier-lab AI security scanners: Claude Security, Codex Security, Big Sleep, CodeMender, AISLE, XBOW compared", "summary": "General-purpose vulnerability scanners from Anthropic, OpenAI and Google, plus independent products such as AISLE and XBOW. They target C, C++, and mainstream application code and have produced CVEs in OpenSSL, OpenSSH, GnuTLS, wolfSSL, SQLite, FFmpeg and V8. They are not cryptography-aware, but cryptographic libraries are written in the languages they scan.", "intro": [ "2026 is the year the frontier labs shipped security scanning as a product: Anthropic's Claude Security moved from research preview (February) to enterprise availability on its Mythos-class models (August), OpenAI's Aardvark became Codex Security (March) and reported 14 CVEs in its first month, and Google paired Big Sleep with CodeMender and a cyber-tuned Gemini model. Independent AISLE, focused on C, claimed all twelve CVEs in a single OpenSSL release.", "For cryptography teams the lesson is mixed. These tools find memory-safety and logic bugs in crypto libraries because those libraries are C and Rust. They do not reason about the mathematics, and zkSecurity's comparisons show plain frontier-model runs missing the critical OpenVM soundness bug that a cryptography-specific harness found. Use them for the implementation layer, not the protocol." ], "choose": [ "Enterprise repositories on Claude or ChatGPT plans: **Claude Security** or **Codex Security**, priced as token usage or subscription; expect a large volume of findings to triage.", "C libraries where memory-safety bugs dominate (TLS stacks, parsers): **AISLE**, or the open-source **AIxCC** systems.", "Web application penetration testing rather than code review: **XBOW**.", "Cryptographic protocol logic: none of these. Pair with a cryptography-specific tool and human cryptographers." ], "page": "https://agentsast.com/categories/frontier/", "tools": [ "claude-security", "codex-security", "big-sleep", "aisle", "xbow" ] }, { "slug": "aixcc", "name": "Cyber reasoning systems (DARPA AIxCC)", "short": "cyber reasoning system", "title": "DARPA AIxCC cyber reasoning systems: Atlantis, Buttercup, RoboDuck, ARTIPHISHELL, open-sourced 2025", "summary": "The seven finalists of DARPA's AI Cyber Challenge, all open-sourced after the August 2025 final. They combine LLMs with fuzzing and program analysis to find and patch bugs in C and Java, processed 54 million lines of code in the final, found 18 real zero-days and patched 43 of 54 synthetic bugs. OpenSSF's OSS-CRS packages them for open-source maintainers.", "intro": [ "AIxCC is the best-documented public experiment in autonomous bug finding: fixed rules, independent scoring, published results and open code. Team Atlanta's Atlantis won with an ensemble of independent bug-finding modules; Trail of Bits' Buttercup placed second using only non-reasoning models at about 181 dollars per point and runs on a laptop; Theori's RoboDuck placed third with an LLM-only pipeline and no fuzzing.", "For cryptography teams the systems are directly usable on C libraries, and Buttercup in particular is cheap to run. They are not cryptography-aware." ], "choose": [ "Cheapest self-hosted starting point: **Buttercup** (Trail of Bits).", "Highest real-world bug count in the final: **Atlantis** (Team Atlanta).", "Study of what LLMs alone can do without fuzzing: **RoboDuck** (Theori).", "Integrated distribution for maintainers: OpenSSF **OSS-CRS**." ], "page": "https://agentsast.com/categories/aixcc/", "tools": [ "atlantis", "buttercup", "roboduck", "oss-crs" ] }, { "slug": "smart-contracts", "name": "Smart-contract AI auditors", "short": "smart-contract scanning", "title": "AI smart-contract auditing tools: Sherlock AI, AuditAgent, Zellic V12, Savant Chat, Olympix, Octane, Hound compared", "summary": "Commercial and open-source AI auditors for Solidity, Vyper, Rust (Solana), Move and Cairo. Published recall against human audits ranges from about 30 percent (Nethermind AuditAgent on its own audits) to about 70 percent on the EVMbench benchmark; precision on live code is around 55 percent in the one controlled study (Sherlock AI). Several also cover ZK circuit languages.", "intro": [ "Smart-contract security has the most data on AI versus human auditors because contests and public audit reports make ground truth available. The picture is consistent: AI tools find a meaningful minority of what human auditors find, with a higher false-positive rate, and the leading vendors position them as a pair auditor run after or alongside manual review. Zellic's co-founder describes V12 as matching junior auditors but not top firms; Nethermind runs AuditAgent as a second layer.", "ZK verifier contracts and bridges are smart contracts, so these tools apply to the on-chain half of a ZK deployment. They do not analyse circuits, with the exception of Savant Chat's listed Circom, Halo2 and Noir support." ], "choose": [ "Published recall on real audits and EVMbench: **AuditAgent** (Nethermind).", "Controlled precision study and PR-integrated workflow: **Sherlock AI**.", "Static-analysis hybrid from a top audit firm: **Zellic V12**.", "Per-commit CI with executable proofs of concept: **Olympix**, **Octane**.", "Open source and language-agnostic: **Hound**.", "Solidity plus ZK circuit languages in one product: **Savant Chat**." ], "page": "https://agentsast.com/categories/smart-contracts/", "tools": [ "sherlock-ai", "auditagent", "zellic-v12", "savant-chat", "olympix", "octane", "hound", "quillshield", "cecuro", "certora-ai-composer", "immunefi-magnus" ] }, { "slug": "benchmarks", "name": "Benchmarks and research", "short": "benchmark", "title": "Benchmarks for AI security tools: zkbugs, EVMbench, ScaBench, SCONE-bench, CyberGym, BountyBench and key papers", "summary": "The datasets used to score AI bug finders, and the research systems that established the methods. zkbugs is the only ZK-specific benchmark; EVMbench (OpenAI and Paradigm) is the most cited for Solidity and the most criticised for contamination; CyberGym and BountyBench cover general software. Read the methodology before the headline number.", "intro": [ "Benchmark numbers drive purchasing decisions, so their construction matters. Detection benchmarks built from public audit reports measure recall of training data as much as discovery; OpenZeppelin's March 2026 audit of EVMbench found invalid high-severity items and contamination risk, and Nethermind noted that recall-only scoring hides false positives. The best benchmarks post-date model training cutoffs, count precision as well as recall, and publish the full findings list.", "The papers listed here are the ones vendors cite: GPTScan for LLM plus static analysis, PropertyGPT for LLM plus formal verification, Hound for knowledge-graph reasoning, and the 2026 cryptanalysis paper that uses agents to generate hypotheses tested by exact computation." ], "choose": [ "ZK circuit tools: score on **zkbugs** in both direct and full-codebase modes and report both.", "Solidity tools: use **EVMbench** with OpenZeppelin's corrections, and **ScaBench** for contest-derived ground truth.", "General code: **CyberGym**, **BountyBench**, **SEC-bench**.", "Any vendor number: ask for the dataset, the cutoff date, precision and the full findings list." ], "page": "https://agentsast.com/categories/benchmarks/", "tools": [ "zkbugs", "evmbench", "scabench", "cybergym", "gptscan-propertygpt" ] } ]