Technical 5 min read

Your AI knowledge base is an attack surface

A knowledge base an AI agent can read and write is a productivity tool, and dated 2026 incidents show it is also somewhere an attacker can plant instructions.

An AI agent that reads and writes a folder of your business’s knowledge is being trusted with three things at once: your private data, whatever untrusted content gets fed into it, and some way of communicating back out. Simon Willison, who has tracked AI security closely, calls that combination the “lethal trifecta”: private data, exposure to untrusted content, and an outbound channel, in one system. Put all three in the same place and the system can be induced to leak what it holds. A knowledge base built from clipped web pages, pasted emails, and forwarded documents is untrusted content by definition. That is not a hypothetical risk. It has already been demonstrated.

What has already happened

Within 48 hours of Anthropic’s Cowork (a desktop agent with folder access) launching in research preview in January 2026, the security firm PromptArmor demonstrated a working attack: a Word document containing invisible one-point white text, which caused the agent to exfiltrate financial documents. The vulnerability was not a bug in one feature. It was the trifecta, played out against a real product within two days of launch.

The clearest numbers come from OpenClaw, the open-source persistent-agent framework whose memory is plain markdown files (SOUL.md, MEMORY.md, USER.md). In its January and February 2026 security crisis, SecurityScorecard’s research, reported by Infosecurity Magazine on 9 February 2026, found 40,214 exposed instances across 28,663 unique IP addresses, with 63 per cent of observed deployments vulnerable and 12,812 exploitable by remote code execution. Beyond the exposure count, the crisis surfaced something specific to this pattern: marketplace skills that wrote persistent backdoor instructions directly into the memory files, instructions that survived even after the skill itself was removed. The memory file was the payload. A poisoned note is not just bad data sitting in a folder; it is an instruction an agent may act on the next time it reads that file.

Why the memory file is the exposure

This is the part of the AI knowledge base pattern that gets the least attention in the enthusiastic write-ups, because it is not where the pattern’s benefits show up. But it follows directly from the architecture. The instruction file that makes the agent useful (what tells it how to behave, what to trust, what to do) is read by the agent every session. If an attacker can get a hostile instruction into any file the agent treats as trustworthy, whether through a poisoned document, a compromised plugin, or a note copied in from somewhere untrusted, that instruction persists exactly as long as any legitimate one does. Deleting the plugin that planted it does not delete what it already wrote into memory.

What the practitioners who take this seriously actually do

The mitigations that show up with real evidence behind them are not exotic. Read-only defaults, so the agent cannot write without deliberate permission. A human gate on anything moving from raw, untrusted capture into the trusted, reviewed layer, the same trust-boundary architecture that shows up across every serious build in this field. Git as a rollback mechanism, so a bad write can be undone rather than just discovered. Secrets never stored in notes. And, the one that matters most given the trifecta: treating anything ingested from outside the business, a forwarded email, a scraped page, a pasted document, as untrusted until a person has reviewed it, not the moment it lands in the folder.

None of this is a reason to avoid the pattern. It is a reason to build it with the trust boundary as a genuine control, not a tidiness convention. A knowledge base with unattended write access and no gate between capture and trusted knowledge is not a small risk sitting alongside a large productivity gain. Until an agent’s write access is scoped and gated, that is the one part of the build worth treating as non-negotiable, whatever else is on the wish list.

Getting that boundary right, not as an afterthought but as part of how the system is designed from the first session, is the kind of judgement call we work through with clients directly in our AI strategy and governance work.

Published 3 July 2026

Perth AI Consulting delivers AI opportunity analysis for small and medium businesses. Start with a conversation.

Prepared by Claude, directed and approved by PAC.

More from Thinking

Adoption 14 min read

How AI capability actually moves through a business

The decisive variable in SME AI adoption is the human absorption sequence, not the tooling. A working framework from observation across WA businesses.

Evaluation 7 min read

AHPRA advertising rules for psychologist websites

Recovery stories, 'specialist', 'clinical psychologist', and endorsement titles are where psychology sites breach the National Law. A practical read-through.

Adoption 6 min read

Customer service AI has finally grown up

Chatbots and AI receptionists earned their bad reputation. What changed, why the trick is in the data, and how the mature version answers every call without replacing anyone.

Evaluation 6 min read

Who can use the titles 'Dr', 'Specialist', and 'Surgeon'?

AHPRA restricts 'specialist' and 'surgeon' to specific registrations, and 'Dr' has its own rule. What health practice websites can and cannot claim.

Adoption 5 min read

Your best people hate writing reports

The operators you promote are brilliant at the work and allergic to reporting. A scheduled AI call interviews them, drafts the briefing, and they approve it. No ego, no politics, no blank page.

Building 6 min read

Your website isn't just for humans anymore

How to build a chatbot that keeps itself up to date, can't leak client information, and won't answer beyond what you've published. The answer was sitting in plain sight.

Evaluation 7 min read

Can you show Google reviews on your health practice website?

AHPRA bans clinical testimonials, even true ones, but service reviews are fine. What that means for the Google reviews widget on your practice site.

Evaluation 7 min read

What AHPRA's advertising rules mean for your website

Your practice website is advertising under the National Law. What AHPRA's rules prohibit, who is responsible, and how to check your own site.

Evaluation 8 min read

Is it safe to paste client data into ChatGPT?

What ChatGPT, Claude, and Copilot promise about your data, what the Privacy Act requires, and the honest answer for professionals handling client files.

Evaluation 6 min read

What a good AI audit actually delivers

An audit is not the report. It is the report plus a working system a client actually keeps, shown through one real engagement, generalised.

Evaluation 7 min read

AI and video, Mid-2026: the models can watch now, not just listen

AI could always transcribe video. It can now read the frames as well, and every hour of footage a business owns becomes something it can question.

Technical 5 min read

Why the privacy case against cloud AI memory isn't paranoia

An AI knowledge base concentrates everything sensitive a business holds. Dated 2026 incidents show what cloud custody means once legal process gets involved.

Adoption 5 min read

The real asset in an AI knowledge base isn't the notes

In every AI-maintained knowledge base, one file carries the owner's judgement and compounds. The wiki pages are the least valuable part.

Evaluation 5 min read

The missing measurement in the AI second-brain boom

Every claim about AI knowledge bases saving time is self-reported. The one controlled experiment measured token economics, not benefit.

Building 11 min read

From evidence base to delivery: a production AI methodology

How we delivered 34 evidence-anchored AI briefings to a WA peer-advisory chapter: fact-checked literature review, multi-agent verification, one method.

Technical 9 min read

The six functions of a working AI system

A working AI system is six functions doing six jobs. When all six connect, hallucinations get caught, outputs hold steady, and models become swappable.

Technical 7 min read

Supervised autonomy: the middle path for AI architecture

Between drafts you approve and agents you hope about sits the middle path: an envelope of authorised routine work, supervised, audited, and yours to widen.

Evaluation 5 min read

The state of applied AI in Mid-2026

Our literature review of applied AI in mid-2026: ten capability categories, three fact-check passes, written for operational leaders.

Evaluation 8 min read

AI in building inspections, Mid-2026

AI defect detection is strong on obvious defects and weak on the subtle ones where liability lives. Which capabilities fit inspection work in 2026.

Evaluation 8 min read

AI in property valuation, Mid-2026

AVMs are reliable enough for triage, not for the final word on contested property. What has shifted in valuation work by mid-2026, and what has not.

Evaluation 8 min read

AI in family law, Mid-2026

Federal Court practice note GPN-AI makes AI verification a professional obligation. What the courts now require, and what the evidence says about legal AI.

Technical 9 min read

How to design a PHI redaction system for clinical AI

PHI redaction is part of a clinical AI tool's architecture, not a feature you add. What the literature says it should look like, and how we built it.

Building 9 min read

How we built on-device de-identification so AI never sees real names

Most AI privacy is a policy. Ours is architecture: an NER model runs in the browser and strips names before anything leaves the device.

Technical 7 min read

Your agency's clients are about to ask why this costs so much

A solo consultant built in three weeks what your agency quoted twelve for. The client doesn't know why yet. The agencies that survive change what they sell.

Adoption 6 min read

What do you love doing? What do you hate doing?

Ask people what they love doing and what they hate doing, then show them AI is coming for the second list. Why the reframe works, and how it fails.

Technical 7 min read

Why I don't use n8n (and what I do instead)

n8n demos well. But a compelling demo and a reliable production system are different things, and the distance between them is where businesses get hurt.

Technical 10 min read

Your codebase was not built for AI. That's the actual problem.

Amazon's mandatory meeting about AI breaking production is an architecture story: codebases built for human maintainers only, now maintained by AI.

Adoption 4 min read

Your team has AI licences. You don't have an AI system.

Fifteen people, fifteen separate AI accounts, no shared context. The problem isn't the tool; it's the architecture around it. Here's the fix.

Building 7 min read

Your $2,000 day starts the night before: our system keeps you on the tools, not on the phone

Optimised routes overnight, automatic customer notifications, and promises the system keeps or corrects. A scheduling system that protects your daily rate.

Evaluation 4 min read

The fastest way for an executive to get across AI

AI moves faster than any executive can track. One focused conversation, one written report, and a decision you can act on: your time stays on the business.

Building 6 min read

Your IT department will take 18 months. You need this working by next quarter.

Senior leaders know what they need built; the gap is time. A prototype gets the tool working now and hands IT a validated blueprint for later.

Adoption 4 min read

What if you had perfect memory across every client?

Every practice captures more than it can recall. AI gives practitioners perfect memory across every client, so preparation becomes thinking time.

Building 8 min read

We built an AI invoice verifier. Here's where it hits a wall.

We built an AI invoice verifier and watched a fake beat a real invoice. Why document analysis alone cannot stop fraud, and the five layers that can.

Building 5 min read

How to build an AI chatbot that doesn't lie to your customers

Woolworths scripted its AI to talk about its mother. The business fix is honesty; the technical fix is architecture that prevents fabrication by design.

Technical 9 min read

Why AI safety features are load-bearing architecture, not political decoration

The 'woke AI' label came from real failures, but they were engineering failures, not safety failures. The difference matters wherever errors have consequences.

Adoption 3 min read

Woolworths' AI told a customer it had a mother. That's a problem.

Woolworths' AI assistant Olive was scripted to talk about its mother and uncle. When callers realised, trust broke instantly. The fix is honesty.

Evaluation 5 min read

Google is no longer the only way your customers find you

Customers now find businesses through ChatGPT, Perplexity, and Gemini. The sites AI cites are structured differently to the sites Google ranks.

Evaluation 4 min read

Two types of AI audit: and how to know which one you need

Where do we start with AI? It depends on whether you need to find the opportunities or reclaim the time. Two audits, two perspectives, one goal.

Evaluation 4 min read

The personal workflow analysis: what watching a real workday reveals about automation

People describe the work they value, not the work that eats their time. Recording a real workday reveals the automation opportunities interviews miss.

Evaluation 4 min read

AI audit that starts with your business

An operations-first AI audit starts with how your business actually runs, and only recommends AI where the evidence says it will work.

Building 6 min read

What production AI teaches you that demos never will

The gap between a demo and a working system is where the useful lessons live. Architecture, framing, privacy, adoption: the patterns repeat every time.

Adoption 6 min read

The psychology of why your team won't use AI

You buy the tool, run the demo, and three months later nobody is using it. Five predictable psychological barriers, each with a strategy that works.

Technical 4 min read

Stop telling AI what NOT to do (and what to say instead)

Instructions built on prohibitions make AI cautious and generic. Describing what you want instead transforms the output, and the reason comes from psychology.

Building 5 min read

How we turned generic AI into a specialist: and what that means for your business

Mediocre AI output is rarely the model's fault. Three structural changes that turn the same model from generic to specialist-grade.

Evaluation 5 min read

Your business has 9 customer touchpoints. AI can fix the 6 you're dropping.

You pay to get customers to your door, then lose them to missed follow-up. AI can handle the six touchpoints most businesses drop.

Technical 5 min read

What happens to your data when you press 'Send' on an AI tool

Businesses send customer data to AI tools without knowing what happens during processing. The spectrum of AI privacy is wider than you think.