Engineering & computing · OECD FoS 2.2 · China MoE 0839 网络空间安全
Cybersecurity
网络安全
L3Systemic transformation ↑ Rising Score 2.80 · Last reviewed 2026-10-21 · Next review 2026-12-20
BIGGEST BREAKTHROUGH 2026-10-16 AI leapField · major
Claude traces the NIST elliptic-curve seeds to drand48 seeded with 1997–98 timestamps; bounty claimed ↗
AI: a 25-year mystery solved by an AI-assisted team, independently re-implemented and verified by bounty organiser Filippo Valsorda; Field: rules out seed-manipulation backdoors (≈30 bits of freedom) without changing curve security
Why this level
AI finds real vulnerabilities at scale: Anthropic validated 500+ high-severity bugs with Claude Opus 4.6 (Feb 2026), Microsoft FORGE contributed to 140 Windows CVEs (May–Sep 2026), TitanCA found 203 zero-days; Google saw the first likely AI-developed zero-day exploit in the wild. Close to level 4.
S = 0.30·method + 0.35·discovery + 0.25·adoption + 0.10·education (education scores 0 when there is no evidence; if empty, the other weights are renormalised)
Sources
- anthropic.com/research/zero-days
- microsoft.com/en-us/security/blog/2026/10/07/3-lessons-from-frontier-ai-vulnerability-research
- arxiv.org/abs/2604.17860
- cloud.google.com/blog/topics/threat-intelligence/ai-vulnerability-exploitation-initial-access
Composite 2.5–3.49 (or ≥3.5 with any of method/discovery/adoption below 3): AI is standard infrastructure with multiple independent breakthroughs. All levels →
Research directions
Vulnerability-hunting agentsAutomated patchingThreat intelligenceDetecting AI-driven attacks
Papers
- TitanCA: Lessons from Orchestrating LLM Agents to Discover 100+ CVEs
203 zero-days, 118 CVEs (preprint)
2026-04-20 · arXiv
- Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models
Adopted in government AI safety evaluations
2024-08-15 · arXiv
- LLM Agents can Autonomously Exploit One-day Vulnerabilities
First systematic evidence LLMs can weaponise CVE descriptions
2024-04-11 · arXiv
- Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents
Shows context compaction in open-source coding agents destroys message provenance, letting injected text masquerade as user instructions.
2026-10-21 · arXiv · Arshan Dabirsiaghi
- Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise
Shows persistent memory turns one-off prompt injection into lasting, cross-session compromise of LLM assistants.
2026-10-20 · arXiv · Pedro H. Barcha Correia, Matheus Azevedo de Sá, Paulo H. Gomes de Senna, Giovanni Mandel Martignago
- A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models
A black-box attack exploiting a shared weakness across LLM watermarks (incl. SynthID), proven optimal across covered families.
2026-10-20 · arXiv · Romina Omidi, Nima Naderloui, Yuan Hong, Binghui Wang
- The Show-Up Attack: When Subagents Win Over Their Supervisor
Malicious subagents manipulate supervisors beyond scope; all six frontier LLMs vulnerable, ASR up to 0.93.
2026-10-20 · arXiv · Maosen Zhang, Jianshuo Dong, Yutong Zhang, Jie Zhang
- Exposing Vulnerability Monoculture Through Systemic Adversarial Robustness
Introduces Systemic Adversarial Robustness to measure how few attacks compromise a whole pool of foundation models.
2026-10-19 · arXiv · Leo Elmecker-Plakolm, Lukasz Szpruch, Avishek Joey Bose, Matthew Wicker
- Towards Architectural-Level Safety Control: Relocating Hazardous Knowledge into Removable Experts in Mixture-of-Experts Models
Relocates hazardous knowledge into removable MoE experts so it can be truly deleted rather than suppressed.
2026-10-19 · arXiv · Cheng Jin, Zhongtian Ma, Xiao Wang, Han Qi
- Peer Pressure: Execution-Verified Co-Evolution of Attacker and Defender Populations for Code Security
Co-trains attacker and defender LLM populations with execution-verified rewards on real repositories.
2026-10-17 · arXiv · Khushbu Pahwa, Yubin Ge, Daniele Bonadiman, Xiaofei Ma
- Narratives Mislead: An Empirical Study of How LLMs Misjudge CVSS Severity in Bug Bounty Reports
Across 4,281 bug-bounty reports, LLM CVSS errors concentrate where narrative impact diverges from formal metric meaning.
2026-10-17 · arXiv · Ruitao Feng, Yuekang Li, Yao Zhang
- From arXiv, bioRxiv, medRxiv and OpenAlex (last ~30 days); summaries written by AIMap from the abstract.
- Last Secure Code Benchmark: Is the Code You Vibe Secure? 2026-10-10
- APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces 2026-10-09
- CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers? 2026-10-06
- Towards Looped Models Done Right, Part II: Rethinking at Fixed Points 2026-10-05
- Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures 2026-09-24
- All related papers in the archive →
Milestones 9· 2 leaps
Big glowing nodes are leaps: AI leap cool · Field leap warm · double ring = both
-
DiscoveryAI leap ⓘ note
Claude traces the NIST elliptic-curve seeds to drand48 seeded with 1997–98 timestamps; bounty claimedAI: a 25-year mystery solved by an AI-assisted team, independently re-implemented and verified by bounty organiser Filippo Valsorda; Field: rules out seed-manipulation backdoors (≈30 bits of freedom) without changing curve security
AnthropicVendor claim -
Industry
Microsoft FORGE: 140 Windows CVEs, May–Sep 2026Vendor: 140 Windows CVEs
Microsoft SecurityVendor claim -
Industry
Google GTIG identifies first likely AI-developed zero-day exploitFirst in-the-wild zero-day exploit likely AI-developed
Google Cloud -
Industry
Anthropic: Claude Opus 4.6 validated 500+ high-severity zero-daysVendor: 500+ validated high-severity zero-days
AnthropicVendor claim -
CompetitionAI leap
DARPA AIxCC final: Team Atlanta wins; autonomous systems find ~80% of planted bugs and patch 60%+, all open-sourcedIndependently scored government challenge shows autonomous find-and-patch
DARPA -
Paper
Google Big Sleep finds a real SQLite vulnerability with an LLMFirst public AI-found exploitable bug in widely used software
Google Project Zero -
Benchmark
Cybench: professional-level CTF benchmark for LM cyber capabilityAdopted in government AI safety evaluations
arXivPreprint -
Paper
LLM agents can autonomously exploit one-day vulnerabilities (87% with GPT-4)First systematic evidence LLMs can weaponise CVE descriptions
arXivPreprint
No milestones match this filter.
More papers & repos 4
| Type | Title | Date |
|---|---|---|
| Paper | TitanCA: Lessons from Orchestrating LLM Agents to Discover 100+ CVEs (arXiv 2604.17860) Preprint | 2026-04-20 |
| Paper | ZeroDayBench (arXiv 2603.02297) PreprintTo be verified | 2026-03 |
| Paper | Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models Preprint | 2024-08-15 |
| Paper | LLM Agents can Autonomously Exploit One-day Vulnerabilities Preprint | 2024-04-11 |
Latest news 12
- Federated Learning for Malware Image Classification under Data HeterogeneityarXiv cs.CRenPaperPreprint
- AdaFedXplain: Adaptive Federated Learning with Explainable AI for Ransomware DetectionarXiv cs.CRenPaperPreprint
- Adaptive Cyber-Defense for Evolving Attacker Capabilities using Reinforcement Learning and Behavior TreesarXiv cs.CRenPaperPreprint
- A Chinese hacker used AI agents to break into more than 100 company websitesTom's Hardware AIen