Version 1.8.5 focuses on controlling what leaves the system and on closing the false-positive and false-negative feedback loop. A new rule-set switch returns opaque responses that strip roughly 21 diagnostic fields, so an attacker cannot use a verbose security reply to map the detectors. Quality Reporting gains a CSV import that turns reviewed false-positive and false-negative rows into resolved cases. The release also widens detection coverage, improves performance, and removes an unused image-processing dependency that carried 74 unpatched CRITICAL and HIGH CVEs.
Vigil Guard PSA has been accepted into the NVIDIA Inception program. We are putting the membership to work on two fronts: integrating Vigil Guard as an external detection layer for NVIDIA NeMo Guardrails, and joining Inception Capital Connect to reach NVIDIA's network of investors as we scale.
Vigil Guard and Oberig IT have signed a distribution agreement that opens up 18 markets across Central and Eastern Europe, the Baltics, the Caucasus, and Central Asia. The idea came straight from customer demand in Oberig IT's markets, and once both sides took a closer look, the potential was obvious. Signing is only the start: work with partners begins right away, and the first PoCs go live within days.
Version 1.8.0 raises detection accuracy on encoded and obfuscated prompt injection, with a new model generation for injection detection, a new embedding engine, and recalibrated signal weights that lift true-positive rates without widening the false-positive envelope. Production scaling is now resource-aware: larger hosts actually serve more traffic, with admission limits derived from host facts instead of one static shape. The release also extends protection to Microsoft Edge and adds two new protected platforms, Microsoft 365 Copilot Chat and Z.ai. It includes one potentially breaking metrics change for operators.
Vigil Guard PSA has joined AI Chamber, the leading business organization for companies working with artificial intelligence across Central and Eastern Europe. The membership extends our presence beyond the Polish market into the wider European ecosystem and gives us a forum to promote the safe and responsible use of AI in business, together with sovereignty over company data.
Most teams shipped AI faster than they secured it. This week Tomasz Bartel, founder and CEO of Vigil Guard PSA, gives two talks on how AI systems get attacked in production and how to defend them: June 9 at Tech Week 2026 (AGH, Krakow) and June 10 at Otwarte Forum.AI (SGH, Warsaw). Real breach cases, where the first signal appeared, and how the control layer responded.
Following LiteLLM's partner acceptance process, Vigil Guard is now built directly into LiteLLM Proxy as an official guardrail provider, alongside established solutions such as Lakera, Aporia, Bedrock Guardrails, and Azure Content Safety. No separate connector, no extra repository. One block in config.yaml screens every model behind your gateway for prompt injection and policy violations in real time.
Claude Code has moved into real developer workflows. It reads secrets, runs shell commands, writes to repositories, and pulls content from the open web straight into the model context. vge-cc-guard is the security sidecar built for exactly this situation.
Vigil Guard PSA is at Infoshare 2026 in GdaΕsk on May 20 and 21. Booth B16, Hall B, Innovation Village. We are showing Vigil Guard Enterprise live: prompt injection detection on real traffic, engine decisions on tool outputs, Scope Drift control, audit log in real time. A few scenarios will be presented publicly for the first time. Come for a direct conversation.
Three changes define the 1.7.x line. The new prompt injection detector vge-promptguard-v2h, tuned to the realities of agentic environments, significantly reduces false alarms on tool outputs and improves detection on hard benchmarks. The first official Vigil Guard integration with Claude Code brings detection to prompts sent in the agentic development environment. An extended API contract exposes fields describing the agent, the tool, and the conversation, providing the foundation for context-aware risk scoring.
Vigil Guard PSA has become a member of #CyberMadeInPoland, Poland's largest organization bringing together cybersecurity companies. Together with the rest of the ecosystem, we are reinforcing the path toward a sovereign cyberspace built on Polish technology.
Four new protection layers built for the realities of agentic AI. Full context logging for every detection decision. Scope drift detection that catches when an agent steps outside its mission. Dual-pass analysis for long prompts. A new injection classifier tuned for code, tool outputs, and structured data.
The detection model powering Vigil Guard Enterprise is now available on Hugging Face. 44M parameters. 0.981 F1 on direct injection detection. 97.5% agentic attack detection rate. Built for systems where AI agents call tools, parse API responses, and execute function returns.
The semantic engine received its most significant overhaul since the platform launched. Detection accuracy improved across all benchmarks with false blocks down 28% and missed attacks down 24%.
The n8n-nodes-vge community node has been verified and published on n8n Cloud. You can now add Vigil Guard AI security directly to your n8n workflows from the canvas.
