← All stories
AI Security

October at Auspere

From Product Demo to Real-World Deployment

Auspere · November 5, 2026 · 5 min read

October was a turning point for Auspere. We opened our beta, welcomed our first users, and spent the month pitching, listening, and learning from the people building and securing AI systems. Here's a look at what we shipped, where we showed up, what we kept thinking about, and where we're headed next.

What we shipped

Auspere opened for beta in October, and both cofounders spent the month pushing the product forward. Everything built follows one principle: every agent action should be evaluated against policy before it happens, and every decision should leave behind evidence that can be independently verified.

  • Decision receipts you can verify yourself. Since late October, every new decision Auspere makes is signed and chained. Anyone can download a receipt and verify it on a public page without logging in. Change a single field, and verification fails.

  • Code Protection. Teams designate repositories, branches, and paths as proprietary. For AI traffic that runs through Auspere, protected code in a prompt or response is matched deterministically and held until someone approves it. Without a verified approval, the request is refused. No model makes that call.

  • Browser protection before information reaches ChatGPT. When policy requires it, our browser extension removes personal information from ChatGPT prompts, so that data never reaches ChatGPT.

  • Investigations that start with the answer. An incident now opens with a summary, related approvals and decisions, a complete timeline, and a direct link to the lineage of the agents involved.

  • A growing integration ecosystem. Connection flows for GitHub, GitLab, and Bitbucket, two-way sync with Asana, and setup forms for 14 action providers, including Jira, ServiceNow, and Splunk.

  • A new view of the threat landscape. The Threats section on auspere.ai brings together an overview, a live feed, and an attack map of activity targeting AI systems.

Every beta implementation still begins with a founder-led onboarding session, so we hear directly from the people using the product and put what we learn back into development.

Getting out of the building

Some of our most valuable progress happened away from our keyboards. We pitched nearly every week.

  • October 7, Upstart Collective hot seat. A 15 minute pitch and Q&A with the 90 Days to Cash Flow Positive cohort, a free cohort Upstart Collective hosts, with direct feedback from founders.

  • October 7, OEN cohort pitch. Five minutes with an Oregon Entrepreneurs Network cohort later that same morning.

  • October 15, Demolicious. We turned our pitch into a game of AI Security bingo. Congratulations to Jack Phan of Oregon AI Accelerator, who took home the win.

  • October 20, AI Portland morning coffee. Conversations with IT and AI leaders at Oregon companies who wanted to see Auspere in action.

  • October 20, OEN Angel Oregon Tech Pitch Day. An in person pitch to Oregon angels that afternoon, alongside founders we learned a lot from.

  • October 22 and 23, AGNTCon + MCPCon. Two days at the Linux Foundation's event in the Bay Area. The sessions were excellent, the event was well run, and De La Soul made an appearance. (Still can't believe I got to see them live. My husband was experiencing serious FOMO.)

AGNTCon was also the most validating stop of the month. A packed session led by OpenAI engineers left attendees with more questions than answers, and a session on keeping agents from going rogue showed how much practical guidance the industry still needs. The companies advancing agent capabilities are not necessarily the companies equipped to independently govern them. That gap is why we built Auspere.

We also welcomed our first beta testers, and every onboarding session is already shaping the product.

What we kept thinking about

October's news was full of agents doing things nobody approved, and every story brought us back to the same point: knowing what an agent did is not the same as proving it was allowed to do it. These ideas shaped our conversations and our product decisions.

  1. Trust is not a control. Safety teams review and warn, but they don't sit in the path of each agent action as it happens. A control does, and it leaves evidence that doesn't depend on anyone's memory.

  2. Transcripts are a log. Receipts are evidence. Keep your transcripts for learning and debugging. But a log is written by the agent and isn't tamper evident. A receipt is made at the moment of decision, by something other than the agent, and can be checked independently.

  3. Sandboxing is not coverage. A virtual machine that holds tells you nothing about what the agent was allowed to reach from inside it.

  4. Your smart fridge has no authorization layer. New research asked whether AI models could copy themselves across cheap connected devices. The lesson for us: every device that can run code needs something deciding what that code may do.

  5. Why we built Threats. Most teams adopting agents don’t fully understand what agents are up against in the wild. We wanted one place that shows it.

What comes next

November is about opening the door wider. A free shadow AI scan is coming to auspere.ai, so any team can see which AI tools are already in use before deciding what to allow, with a self-serve page for the browser extension close behind. Our goal is 100 beta teams by the end of the year.

If your organization is running agents, or suspects it is, let's talk. Every beta team gets founder-led onboarding, support during business hours, and a seat in our Discord community.

Reach out through auspere.ai and tell us what you're seeing.

L.R.

Keep exploring ideas for a more accountable AI future.

Back to the journal →