Add selected Workbench capabilities through bounded OEM and partner integrations
Finding
3,411 of 10,152 unique arXiv papers in the current pull are strict AI-security papers, already naming concrete work in prompt/generation security, agentic action security, model and ML attack security, governance assurance, privacy protection, MCP/tool-use security, and red teaming.
Finding
The research frontier already has a name for every problem the hiring market is still describing informally.
A paper-level frontier snapshot, not a time-series lag study or hiring benchmark.
Connected intelligence
A three-layer framework intercepting direct and indirect prompt injection throughout the inference pipeline, reducing Attack Success Rate from 71.4% to 11.3% across GPT-4o, Llama 3, and Mistral 7B.
The first large-scale, production-proven enterprise framework for securing AI agents operating through the Model Context Protocol, deployed at Uber for over ten months across more than 7,200 unique hosts.
3,411
AI-security-tagged papers
AI-security-tagged papers in the State 2026 research corpus.
A runtime security framework enforcing a user-confirmed rule set at every tool-call boundary, blocking indirect prompt injection pathways without model modification, evaluated across five state-of-the-art language models.
Reveals that the reasoning capabilities enabling LLM-based guardrails introduce a novel vulnerability: attackers can inject crafted data to trap the guardrail in extended reasoning loops, effectuating a denial-of-service attack.
Conducts STRIDE and DREAD threat modeling of MCP implementations across five key components, finding tool poisoning -- malicious instructions embedded in tool metadata -- as the most prevalent and impactful client-side vulnerability.
Presents AI-Infra-Guard, an open-source framework organizing AI red teaming across infrastructure, protocol/tool, agent behavior, and model layers, spanning deterministic rule matching, LLM-driven auditing, and jailbreak harnesses.
Systematizes trust boundaries and security risks of agentic LLM-based systems, mapping a taxonomy of attacks spanning prompt-level injections, knowledge-base poisoning, tool/plug-in exploits, and multi-agent emergent threats.
A fine-grained fuzzing framework that systematically exposes vulnerabilities in chat templates, a critical yet underexplored attack surface, achieving a 98.2% average jailbreak attack success rate across twelve open-source LLMs.