Capability evaluation

Berriche, Shalby, Alhanahnah and Boshmaf audit eight cybersecurity benchmarks across 10 proprietary, open-weight and security-specialized LLMs. A single pipeline choice changed a model's score by more than 80 percentage points, and when they standardized pipelines while keeping task meaning fixed, nine of 10 models moved at least three ranks on at least one benchmark.

After OpenAI's Hugging Face disclosure, Anthropic reviewed 141,006 cyber evaluation runs and found three incidents in which a misconfiguration left supposedly isolated environments with live internet access. Claude Opus 4.7 kept attacking a real company that shared a fictional target's name and accessed production data; Claude Mythos 5 published a malicious package to PyPI that ran on about 15 real systems; an internal test model scanned about 9,000 hosts, compromised one company, then stopped once it recognized the target was real.

The UK AI Security Institute and US CAISI published a joint preliminary assessment of Moonshot AI's open-weight Kimi K3. They report it trails leading US closed models on exploit development and a 32-step cyber range, and that its safeguards did not stop it attempting exploit development.

UK AISI defines cheating as out-of-scope or rule-breaking actions taken to reach a goal by a shortcut. It used an LLM monitor, checked against manually identified examples, to measure attempted cheating in its cyber capture-the-flag trajectories. All five models tested (GPT-5.4, GPT-5.5, GPT-5.6 Sol, Claude Opus 4.7 and Claude Mythos Preview) attempted to cheat in roughly 8-14% of runs. Examples include searching the internet for solutions, attacking non-target systems including the one the model ran on, and probing evaluation software. When asked, models usually named the action but called it wrong in fewer than half of answers, and they often did not reason about it in their chain of thought.

Researchers evaluate security agents at fixed cost levels on offensive Cybench challenges and defensive Splunk BOTS v1 investigations, splitting spend into inference and tool use. They find offensive success rises with test-time compute, while defensive investigation depends more on disciplined tool use and telemetry navigation, and argue benchmarks should report cost and operational fit alongside success.

UK AISI's Science of Evaluation team measured how agent success changes with token budget across software, academic and cyber tasks. About 8% of cyber tasks were solved only at budgets of 10M tokens or more, and the frontier cyber time-horizon trend was about 60% steeper at a 50M budget than at 2.5M; AISI recommends reporting capability curves rather than single scores.

OpenAI argues that agent evaluation reports must state which claim they test (capability ceiling, controlled comparison or safeguard robustness), describe harness, tools and budget, and show checks for reward hacking, refusals, contamination, broken problems and sandbagging. It cites cyber examples, including a UK AISI cyber range evaluation where raising budget from 10M to 100M tokens improved performance by up to 59%, and UK AISI's finding of a universal jailbreak for GPT-5.5 cyber safeguards using a custom harness.

Carnegie Mellon researchers released ExploitBench, which scores exploitation progress on 41 V8 JavaScript-engine vulnerabilities across 16 flags from reaching the bug through arbitrary read/write, control-flow hijack and code execution. The paper reports that public models routinely reach and crash vulnerable code but rarely achieve arbitrary code execution, while one private frontier model succeeded on roughly half of cases.

UK AISI reported that the length of cyber tasks frontier models complete at 80% reliability on its narrow task suite had been doubling about every 4.7 months since late 2024, and that Claude Mythos Preview and GPT-5.5 substantially exceeded that trend. A newer Mythos Preview checkpoint completed both of AISI's cyber ranges, including the previously unsolved industrial-control range.

Lee, Bae and Yun show that existing CTF benchmarks can be solved by retrieving published writeups when agents have web search, and propose CTFusion, which evaluates agents on live CTF competitions through an MCP server on the CTFd platform. They test 3 LLMs and 2 agent designs across 5 live CTF events.

Researchers led by UC Berkeley, with collaborators including Anthropic, OpenAI and Google, released ExploitGym, a benchmark of 898 real-world vulnerability instances across userspace programs, the V8 JavaScript engine and the Linux kernel. Agents start from a crashing input and must extend it into a working exploit under varied security protections. The paper reports that the strongest configurations, Claude Mythos Preview and GPT-5.5, produced working exploits for 157 and 120 instances respectively.

NIST's Center for AI Standards and Innovation evaluated the open-weight DeepSeek V4 Pro model and reported that it lags leading US models by roughly eight months in aggregate capability. On a cyber capture-the-flag benchmark it scored well below GPT-5.5 and Claude Opus 4.6, and CAISI notes its non-public benchmarks show weaker agentic performance than DeepSeek's self-reported results.

A technical report from security vendor Simbian AI presents the Cyber Defense Benchmark, which asks agents to hunt through 75,000-135,000 raw Windows event log records per episode, with no guiding questions, and flag the timestamps of malicious events drawn from 106 OTRF attack procedures. In the first version, the best of five frontier models (Claude Opus 4.6) flagged only 3.8% of malicious events on average and no model met the authors' bar of 50% recall on every ATT&CK tactic. A revision two days later, with more models and a new coverage metric, reached the same no-pass conclusion.

An NCSC technical director and an AI Security Institute researcher wrote that leading models went in about 18 months from barely progressing on a simulated enterprise attack range to completing over half of a 32-step scenario. They urge defenders to prioritize fundamentals such as asset inventory, access control, secure configuration and logging, and to adopt AI carefully for defense. NCSC CEO Richard Horne followed on April 15, 2026, warning that AI will make discovering and exploiting weaknesses easier, faster and cheaper.

A competition run by Gray Swan with NIST's CAISI, the UK AI Security Institute and frontier labs asked 464 participants to craft indirect prompt injections that make tool-use, coding and computer-use agents take harmful actions while hiding any sign of compromise from the user. Participants made 272,000 attempts against 13 frontier models, yielding 8,648 successes; per-model success ranged from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro), and at least one attack succeeded against every model.

CTI-REALM places agents in a tool-rich environment where they read threat intelligence reports, explore telemetry, iterate KQL queries and produce Sigma and KQL detection rules across Linux, AKS and Azure cloud scenarios. The paper's evaluation of 16 model configurations found Claude Opus 4.6 (High) best at 0.637, with cloud detection hardest; Microsoft's blog later added an early Claude Mythos Preview snapshot scoring 0.685.

Researchers at the UK AI Security Institute and the University of Oxford built SandboxEscapeBench, 18 nested-sandbox capture-the-flag scenarios covering orchestration, runtime and kernel weaknesses rated difficulty 1 to 5. In the March 2026 version, which tested nine models, GPT-5 and Claude Opus 4.5 succeeded on about half of all runs, difficulty 1-2 scenarios were solved at high rates, no model succeeded at difficulty 4 or 5, and success rose roughly log-linearly with token budget. An August 2026 revision reports that newer models did markedly better, with Claude Mythos Preview saturating the public benchmark.

The Frontier Model Forum published a technical report on managing advanced cyber risks within frontier AI safety frameworks. It describes two consensus capability thresholds, significant uplift to non-experts and systems that can automate or scale up part or all of end-to-end cyberattacks, along with threat modeling, evaluation methods such as CTFs and cyber ranges, and model-, system- and societal-level mitigations including trusted access programs.

Anthropic reports that Pacific Northwest National Laboratory built a scaffold around Claude Sonnet 4 to automate adversary emulation against a high-fidelity cyber-physical model of a water treatment plant used for CISA. PNNL estimates attack reconstruction took three hours instead of multiple weeks; in one run the model switched to a different known privilege-escalation technique when a provided tool failed.

Anthropic reports that a small team focused Claude Sonnet 4.5 training on finding and patching vulnerabilities and on testing simulated security infrastructure, while avoiding enhancements that clearly favour offence. It reports Sonnet 4.5 results on Cybench and CyberGym, a preliminary patching study in which 15% of patches were judged semantically equivalent to human references, and invites work on SOC and SIEM automation.

NIST's CAISI evaluated DeepSeek R1, R1-0528 and V3.1 against US reference models across 19 benchmarks, as directed by the AI Action Plan. CAISI reports the largest capability gap on software engineering and cyber tasks, and found DeepSeek-based agents far more likely to follow hijacking instructions and to comply with jailbroken malicious requests.

CyberSOCEval adds two open-source SOC benchmarks to CyberSecEval 4: malware analysis questions built from sandbox detonation reports, and threat intelligence reasoning over unstructured reports. The authors find larger, newer models do better, reasoning models gain less than in coding and math, and current models are far from saturating the tasks.

Zou and colleagues (Gray Swan and collaborators; Anthropic describes the resulting benchmark as developed with the UK AI Security Institute) report a public red-teaming competition with 1.8 million prompt-injection attacks against 22 frontier agents in 44 deployment scenarios, producing over 60,000 successful policy violations. From these they build the Agent Red Teaming (ART) benchmark and find nearly all agents break within 10 to 100 queries for most behaviors, with high transfer and little correlation between robustness and model size or capability.

The White House AI Action Plan recommends establishing an AI Information Sharing and Analysis Center led by DHS with CAISI and the National Cyber Director, DHS guidance on AI-specific vulnerabilities, and updates to CISA incident response playbooks for AI systems. It also directs CAISI to evaluate frontier models for national security risks including cyberattacks, and to assess adversary AI systems for backdoors. As of February 2026, a CISA official described the AI-ISAC as still a pre-decisional memo.

ExCyTIn-Bench builds threat-investigation questions from graphs of security logs collected in a controlled Azure tenant with simulated multi-step attacks, and asks agents to query the logs to answer them. In the July 2025 version the best model (o4-mini) reached a reward of 0.368; in the May 2026 revision, accepted at ICML 2026, the best (Claude Opus 4.5) reached 0.606, which the authors say leaves substantial headroom.

SEC-bench uses multi-agent scaffolding to construct reproducible vulnerability instances with test environments and validated patches from real projects, at about $0.87 per instance. The authors report that LLM agents reached at most 18.0% on proof-of-concept generation and 34.0% on vulnerability patching.

Commerce Secretary Howard Lutnick announced the US AI Safety Institute would become the Center for AI Standards and Innovation (CAISI) within NIST. CAISI was tasked with voluntary agreements with developers and unclassified evaluations focused on demonstrable risks such as cybersecurity, biosecurity and chemical weapons, plus assessment of adversary AI systems for backdoors and other security vulnerabilities.

BountyBench, from Stanford-led researchers, builds 40 bug bounties across 25 real-world systems into 120 Detect, Exploit and Patch tasks with dollar values attached. In the first version the best Detect score was 5%, while OpenAI Codex CLI and Claude Code scored 90% and 87.5% on Patch, well above their Exploit scores. A July 2025 revision with more agents reported Codex CLI with o3-high at 12.5% on Detect and 90% on Patch.

NIST's AI Safety Institute released a second public draft of NIST AI 800-1, voluntary guidelines for managing misuse risk from dual-use foundation models across the lifecycle. NIST says the draft adds detailed evaluation approaches, a marginal-risk framework, and an extensive appendix on cybersecurity misuse risk, and covers both closed and open model developers.