{"canonical_url":"https://engineer.company/ai-security/","content_markdown":"# Security for the AI agents you run\n\nUnderstand what your AI agents can reach and put controls around that access.\nAn agent connected to mail, documents or an API introduces permissions and\nfailure paths that need review. We inventory the agreed systems, assess the\nrisks, implement approved controls and test the restricted actions. You receive\nthe access map, findings and verification record.\n\n## Mapped: every agent and what it can reach\n\nWe inventory agents within the agreed scope, including approved integrations\nand other connections identified with your staff. For each, the map records\ncredentials, tools and reachable data. You receive the map with the findings\nand responsibility for keeping it updated is agreed as part of handover. It is\na record of the reviewed environment, with its scope and gaps made explicit.\n\n## Locked: each agent on least privilege\n\nAn agent gets its own identity and the smallest privilege\nits job needs: roles rather than shared accounts ([RBAC](/glossary/#rbac)),\nread-only by default, writes only through narrow tools with limits on how often\nand how much. [Secrets](/glossary/#secret) are split per agent and per host,\nnever pasted into a prompt, and rotated without a code change. External text an agent reads — an email, a web page or a document — is treated as untrusted data. Tool permissions and validation checks enforce the agreed boundaries; a message saying \"ignore your rules\" is not authorization for an action.\nControls fail closed: when a check cannot run, the action is refused.\n\n## Recorded: what it did, and for whom\n\nEvery tool call is [logged](/glossary/#logging) with the person it acted for,\nwhat it was asked, what it touched and what came back, and the log is kept\nwhere the agent itself cannot change it. [Alert rules](/glossary/#alert) are agreed for the system. They can cover a burst of reads, use outside expected hours or access beyond the agent’s agreed data scope. When something does go wrong, the record is\nwhat turns an [incident](/glossary/#incident) into a timeline rather than a\nguess.\n\n## A lock is tested, not trusted\n\nWe test the access controls within the engagement's scope and record what was\nattempted, refused or still needs attention. Tests can cover denied tool calls,\nrate limits, credential isolation and prompt-injection attempts. The result is\na reviewable verification record and a set of repeatable checks, with their\nlimits stated. Incident work and credential rotation are scoped separately\nwhen an existing exposure needs containment.\n\n## The work behind it\n\nWe have done this work on servers, identity systems and platforms. We\n[wrote 1,000 RBAC rules across Linux and database servers](https://engineer.company/portfolio/wrote-1000-rbac-rules-to-secure-linux-servers/),\n[led access management across M365 and single sign-on](https://engineer.company/portfolio/led-access-management-across-m365-and-sso/),\nand [ran a continuous security scanning programme with 21 written audits](https://engineer.company/portfolio/ran-a-continuous-security-scanning-programme/).\nWe [split every operational secret per host](https://engineer.company/portfolio/split-every-operational-secret-per-host/),\n[added a one-second preflight for secrets](https://engineer.company/portfolio/added-a-one-second-preflight-for-secrets/),\n[found three locks on the door, none of them locked](https://engineer.company/portfolio/three-locks-on-the-door-none-of-them-locked/),\nand [proved the intrusion ban works on every run](https://engineer.company/portfolio/proved-the-intrusion-ban-works-every-run/).\nWe [built fail-closed rate limiting](https://engineer.company/portfolio/built-fail-closed-rate-limiting-and-anti-abuse/)\nand [wrote a scope rule after a leak](https://engineer.company/portfolio/wrote-a-scope-rule-after-a-credential-leak/).\nHow we keep our own house is on [how we work securely](/security/).\n\n## How an engagement runs\n\nIt starts read-only. We map the agents in the agreed scope, their credentials and what they reach,\nand hand you the findings in order of risk, each with its reason and the fix.\nThe fixes are then made in your configuration and your repository, through your\nreview, with checks that fail when covered access paths exceed the agreed permissions. You\nkeep the map, the record and the tests — the file you hand to an auditor or an\ninsurer — and the [data agents read](/ai-data/) can be permissioned the same\nway. It is bought like the rest of our work, on the terms in\n[what it costs](/what-it-costs/).\n\nIf agents already hold keys to your systems,\n[tell us which ones](/contact/) *(read by the engineer)*.\n\n\u003chttps://engineer.company/ai-security/\u003e\n","date_modified":"2026-11-12T05:34:23Z","description":"Access reviews and controls for AI agents: an agreed inventory, prioritized findings, permission tests and a verification record.","formats":{"gemtext":"https://brand.engineer.company/ai-security/index.gmi","gopher":"https://brand.engineer.company/ai-security/gophermap","html":"https://brand.engineer.company/ai-security/","json":"https://brand.engineer.company/ai-security/index.json","markdown":"https://brand.engineer.company/ai-security/index.md","pdf":"https://brand.engineer.company/ai-security/index.pdf"},"kind":"page","language":"en","title":"Security for the AI agents you run","url":"https://brand.engineer.company/ai-security/","version":1}