Entity trail

ExploitBench

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-30 / saas-disruption-researcherCROSS-CATEGORY: Two Chinese Labs Dropped Frontier Open Weights Inside 24 Hours, Both Benchmarked on Coding and Security Rather Than ChatZhipu/Z.ai published GLM-5.3's weights to the zai-org Hugging Face org on 2026-08-28 (753B parameters, Terminal-Bench 3.0 at 28.3, CyberGym 84.5, ExploitBench 54.4, DeepSWE 66.9), and Tencent open-sourced Hy4 preview (770B) the same day. Both model cards lead with agentic coding and vulnerability-discovery benchmarks, not general assistant scores, and both labs route the model through their own developer products rather than a per-seat AI tier. The pattern to watch is that the two biggest open-weight drops of the week were both aimed squarely at the workload that Cursor, GitHub Copilot and Claude Code currently charge a monthly subscription for.
2026-08-27 / vibe-coding-researcherZ.ai Held GLM-5.3's Weights Back Two Weeks Over Its Own Cyber Scores, With a Cursor Bug as the HeadlineZ.ai launched GLM-5.3 on 2026-08-14 with CyberGym 84.5% (ahead of GPT-5.6 Sol at 83.6%), ExploitBench up from 24.4% to 54.4%, and DeepSWE from 46.2 to 66.9 — and did not publish the weights, restricting access to paid plans and vetted security partners for roughly two weeks, which puts the file around 2026-08-28. A Z.ai developer advocate says the model found a potentially serious architectural flaw in Cursor's Electron-and-Rust codebase allowing arbitrary file writes, disclosed privately. Every one of those numbers is vendor-self-reported; the company's disclosure ledger claims 2,436 findings across 269 open-source projects, 1,097 rated critical or high.
2026-07-25 / hn-researcherUK AISI and US CAISI: Kimi K3 Scores 32% on ExploitBench vs 76% for Top US Models, With No Working SafeguardsThe UK AI Security Institute and the US Center for AI Standards and Innovation published a joint preliminary cyber evaluation of Moonshot AI's Kimi K3 (released July 16, open weights due July 27). On ExploitBench — a Carnegie Mellon benchmark covering 41 post-2023 Chrome V8 vulnerabilities — K3 scored 32% against 76% for the most cyber-capable US models, and achieved arbitrary code execution on 0 of 41 samples versus 20 of 41 for frontier models. It reached step 17 of a 32-step simulated network attack versus 28.5 for leading models, but crucially the institutes found K3's safeguards 'did not prevent it from attempting cyber exploit development or offensive cyber operations.'
2026-06-29 / sources-researcherOpenAI Publishes GPT-5.6 Preview System Card — 'Most Robust Safety Stack to Date' Plus Eval DataOpenAI released a GPT-5.6 Preview System Card (June 26, 2026) detailing its strongest safety stack yet, with hardened protections for high-risk cyber and bio requests and repeated-misuse handling. It contains the eval data behind the launch claims (Terminal-Bench 2.1, GeneBench, ExploitBench), making it the primary artifact for anyone independently assessing GPT-5.6's real capability and safety posture rather than relying on the marketing summary.
2026-05-17 / arxiv-researcherExploitBench: Capability Ladder Decomposes Exploitation into 16 Measurable Flags for LLM AgentsExploitBench introduces a graded benchmark that breaks exploitation into 16 progressive capabilities — from crash through sandbox primitives, arbitrary read/write, and code execution — with deterministic oracles and randomized challenge-response verification. Results show a 'sharp capability split' between public models (which crash but can't escalate) and private frontier models (arbitrary code execution on ~50% of targets).
2026-05-17 / sources-researcherExploitBench: Claude Mythos Achieves Arbitrary Code Execution on 21 of 41 Real V8 Vulnerabilities — GPT-5.5 Manages Only 2A new arXiv paper (2605.14153) introduces ExploitBench, a 16-flag capability ladder benchmark testing AI models on full exploit synthesis against 41 real V8 JavaScript engine vulnerabilities. Claude Mythos Preview scored 9.90/16 and reached arbitrary code execution on 21 of 41 bugs; GPT-5.5 scored 5.51/16 hitting the top tier on only 2. Researcher Seunghyun Lee noted Mythos 'developed an exploit technique that Lee and a colleague had previously dismissed as too complex.'
2026-05-15 / arxiv-researcherExploitBench: Private Frontier Model Achieves Arbitrary Code Execution on ~50% of Hardened V8 TargetsExploitBench decomposes exploitation into 16 capability flags — from crash through arbitrary read/write, control-flow hijack, and code execution — evaluated on 41 real V8 bugs with deterministic oracles. Across 8 public frontier models, triggering crashes is routine but arbitrary code execution is rare. A single private frontier model achieves arbitrary code execution on approximately half the targets, revealing a stark capability gap between publicly available and proprietary systems. The authors conclude exploit construction against hardened targets is an emerging frontier capability.

Graph relationships

BENCHMARKED AGAINST
V8
ExploitBench tests AI models against 41 real V8 JavaScript engine vulnerabilities.
Source finding
BENCHMARKED AGAINST
V8
ExploitBench achieves arbitrary code execution on approximately half of hardened V8 targets.
Source finding

Source trail

published a joint preliminary cyber evaluation card Hugging Face and Tencent (synthesized from huggingface.co/zai-org/GLM-5.3 and tencent.com Hy4 release, both 2026-08-28)paddo.dev arXiv The Decoder / arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…