Reconsider legacy tech and remote-access security in the AI age. Corelight, Dropzone AI, Scythe, SimSpace, and Sondera launch AI proving ground. 24th July 2026 Five AI companies are teaming up to boost enterprise trust in AI agents via a new consortium. Secops-focused Dropzone AI said the new AI Proving Ground Consortium (AIPGC) will rigorously train, test, and prove agentic AI defences in production-like environments. Edward Wu, founder and CEO of Dropzone AI, said enterprises need to know they can trust agentic AI. As AI takes on a bigger role, organisations need confidence on how these systems perform in real environments. "The future of the SOC is multiple AI agents working together, alongside human analysts, across investigations, threat hunting, intelligence and response," Wu said. "This consortium can help establish the benchmarks and standards organisations need to confidently adopt AI and accelerate its full operational impact on security." As a result, the first virtual event from the venture between founders Dropzone AI, Corelight, Scythe, SimSpace and Sondera was held mid-June. Peter Lee, CEO of testing platform SimSpace, the lead organiser, said AI weaponisation is a bigger problem than any organisation can handle alone. "That's why we're forming the AIPGC to pool our collective ideas and cyber expertise," Lee confirmed. "AI is transforming cybersecurity. Organisations need a way to rigorously train, test, and validate AI agents together with human cyber operators." AI agents in secops and services. Dropzone AI's agentic secops offering targets managed security services providers (MSSPs). As a result, it has rolled out AI SOC Analyst, enabling autonomous investigation of more alerts and AI Threat Hunter, which runs prebuilt or custom 'hunt' packs for SIEM, EDR, and cloud. In addition, it offers AI Threat Intel Analyst, which analyses and designs responses for the AI Threat Hunter agent. "SimSpace research found that nearly 80% of security leaders report high confidence in their AI defences," the announcement said. "Measured readiness scores can be as low as 30% before repeated simulation exercises." Additionally, many organisations rely on legacy preparation strategies, like tabletop exercises and certification courses, it said.

Sondera, a startup developing formal verification controls for AI agents, has announced its research on compiling natural-language policy into verified rules has been accepted at workshops at ICML 2026 and the Federated Logic Conference's LLM-Solve workshop. A related tool will be demonstrated at Black Hat Arsenal. The company's neurosymbolic approach uses neural classifiers to evaluate agent actions whilst symbolic rules enforce policy deterministically. Using MedAgentBench, Sondera's pipeline autoformalized 23 of 88 policy rules and blocked all 99 adversarial unsafe attempts in peer-reviewed research. Sondera's system compiles natural-language policies into formally verified Cedar policy-as-code, which runs outside the agent's context window to prevent circumvention through prompt injection or model drift. The policy enforcement is stateful, tracking agent behaviour across full runs. The company is currently in private beta.