AI agents: Software engineer report

Nicholas (Niccolo) Roberts · Sovereign Source Super Intelligence

Expanded web report · Related PDF snapshot: 22 October 2026 · PDF page 7

Research reviewed 22 October 2026 · Evidence and editorial planning guidance are distinguished below.

Produce evidence that the control works

Enforce authorisation where the action executes. Treat model output and tool arguments as untrusted input, validate object ownership and apply limits server-side. Keep idempotency and audit correlation for actions that affect money or customer records.

Your immediate priorities

Test direct API calls that attempt to bypass the conversational approval flow.
Exercise bounded retries, duplicate requests and revocation with synthetic accounts.

How to work through the decision

Work in the approved environment and keep secrets out of tickets, logs and chat. Record the artifact or configuration being tested, the expected behavior and the observed result. Test rejection of unauthorized access alongside successful business behavior. Preserve enough evidence for another person to review the conclusion, while minimizing sensitive data. A passing smoke test is one observation, not proof that every exposure path has been eliminated.

Evidence to ask for

A set of negative tests proving the agent cannot cross tenant, spending or approval boundaries.

Does the backend reject an unauthorised action even if the model confidently requests it?

Work with the next person

Return a reproducible verification record to the technical manager, including untested paths and assumptions. Identify the operational owner of any follow-up test that cannot run during the immediate change window.

Continue with the technical manager report for this topic.

A scenario to rehearse

Illustrative exercise, not an observed customer incident: A support agent can draft a refund explanation and also call the refund API. The exercise tests whether a persuasive conversation can cause payment without the separate approval required by the business.

What a useful outcome looks like

A set of negative tests proving the agent cannot cross tenant, spending or approval boundaries. Keep the decision dated and return to it if new evidence changes the scope.

Sources and further reading

Microsoft: Securing the agentic insider

What does this mean for your business?

Use the Cyber Health Review to turn the research into priorities for your own operating model.

Contact help@cyberpolycrisis.com

A planning tool.
A starting point for evidence.

Based on the Cyberpolycrisis articles by Nicholas Roberts. The articles examine shared dependencies across concurrent events; they do not establish a single coordinated campaign.

How is readiness calculated?

Each applicable control has equal weight: established and tested = 100, partly in place = 50, missing or untested = 0, and not sure = 0 pending verification. The mean is rounded to a whole number. AI controls are excluded when you report no AI use.

The planning bands are 0–49 “Stabilize”, 50–79 “Close gaps”, and 80–100 “Close remaining gaps” or “Validate and maintain” when none remain. These are editorial priorities, not validated risk thresholds. A current incident concern overrides the band. A high average can still contain a consequential individual gap.

This is a self-reported checklist. It does not measure the probability of a breach, scan your systems or certify your security.

What does the network map show?

Connections reflect thematic overlap in the published network analysis. Highlighted connections mean your answers show gaps in both areas. They do not demonstrate a live attack path, causal link or measured likelihood. Line thickness is illustrative.

Why these platforms?

The shortlist matches your business needs to documented capabilities. Shopify can own commerce records; Palantir can support governed operational workflows; Databricks can govern analytics and AI data; Cloudflare can protect access and applications and host suitable workloads.

“Safe haven” means a better-controlled operating model, not a guarantee from a vendor. Documentation alone does not establish a comparative security track record. Before buying, review current assurance reports, relevant incident disclosures, contract terms, data location, implementation cost and exit options.

Vendor controls depend on product, plan and configuration. Your team remains responsible for access, integrations, data and recovery.

Where do my answers go?

Scoring runs in your browser. Saving progress is optional and uses this browser’s local storage. Anyone with access to the browser profile could read saved answers. Your completed review is sent to the chatbot only when you explicitly select “Include my completed review with the next question.”

Chat sends your message and recent conversation to Cloudflare Workers AI. This app keeps the conversation in page memory and does not save transcripts to a database. “New chat” clears that local conversation; it cannot recall data already sent. Hosting and AI providers process requests under their own terms. Do not share passwords, API keys, customer data or confidential incident details.

If you choose “Email me a follow-up,” we store your email address, consent version and request timestamps privately in Cloudflare for Nicholas Roberts to respond. Your chat and assessment aren’t included in this request, and the email address isn’t sent to the AI. This is a personal follow-up request, not a newsletter subscription. You can ask Nicholas to remove your request through his LinkedIn profile.

Downloads stay on your device. The contact button opens a generic email draft and does not include your answers. Hosting providers still process ordinary web requests; do not put assessment information in URLs.

Sources and version

Methodology 1.0. Capability documentation reviewed October 2026. Review current vendor terms and controls before making a decision.

JavaScript is needed to calculate your plan locally. You can still download the free report, read the methodology and explore the Cyberpolycrisis network analysis.