Watched around the clock, changed only with approval: the Cybersecurity AI Agent

Watch scope on our own site: DNS, SSL certificates, email setup, & hosting infrastructure, plus the live pages themselves. Team model: the AI agent works under the same operating model as every Thanos Labs AI agent, & its tier carries the scrutiny that tier requires: approval gates on every consequential action, plus the full test battery before launch & on major changes.

The watchdog layer. A website never closes, so the check cannot either. Around the clock, the production watchdog checks the live site for dead links, key pages returning errors instead of content, & the homepage serving the site rather than an error page. If everything looks right, it stays quiet. The watchdog is production health monitoring, not a vulnerability scanner, & we say so plainly wherever it is described: it watches for the site itself changing in ways it should not, between the scheduled scans.

The scanning layer. Security scans run on a schedule, & their raw output is only the start. The AI agent interprets what the scanners find, separates real threats from noise, & matches every finding against the client's actual resources & software versions. If the vulnerable software or version is not in the client stack, it does not alert. Intelligence comes from public sources, including the CISA Known Exploited Vulnerabilities catalog, the National Vulnerability Database, OSV, & EPSS scoring, with OWASP & MITRE ATT&CK as the reference frameworks for fixes & attacker behavior.

Sources: The Watchdog Is Watching, October 11, 2026; Cybersecurity AI Agent build spec, October 11, 2026 (canonical resources and bot behaviors).

• 1. Every finding carries severity & an ownerFindings are reported with what was observed, why it matters, and the recommended remediation. A bare alert with no evidence and no fix is treated as noise, and the AI agent does not emit noise.
• 2. Critical & high findings block a releaseThey are resolved before the release moves, not noted and waved through. Critical findings are escalated to a human straight away.
• 3. Nothing is auto applied or auto fixedEvery fix moves through preview, a written changelog, and approval, then publish. A human approves every fix before anything changes.
• 4. Service interrupting changes need moreApproval plus an audit log entry plus a rollback plan, requested before the change, not reconstructed after it.
• 5. Credentials never travel in the openAccess is provided through secure links into secure storage. The team never sees client passwords and never stores them where anyone can read them.

Two standing rules sharpen the watch. On every scan, missing two factor authentication is flagged as a high severity finding, with no exceptions. & access is least privilege by default: the AI agent takes only the permissions the work needs, client data stays isolated per client with defined retention periods, & the audit log is one the client can inspect.

Sources: Never Hacked: How We Protect You, October 2026; Cybersecurity AI Agent external audit checklist, October 11, 2026.

The AI agent is audited against an external audit checklist, the same document we hand to any outside auditor. It covers six areas: authority & boundaries, decision making transparency, human oversight, data & privacy, testing & reliability, and operations, down to audit logging that traces every action to a timestamp, an input, the rule that fired, & the approver.

The checklist ends with three non negotiables, the questions an auditor asks before anything else:

• Show the actual rulesThe decision logic itself, not a marketing description of it.
• Show the test resultsMeasured performance, not claimed performance.
• Show the logsWhat the AI agent actually did, not what it was supposed to do.

An AI agent that cannot answer those three does not get trusted with a client system. Ours is built to answer them, & its known limitations are documented as part of the standard: where evidence is thin or conflicting, the AI agent says so, & an honest not sure is a valid output.

• It does not act alone on anything consequential.In security work the AI agent collects read only evidence and proposes. A human approves actions such as disabling an account or blocking an address.
• It does not scan beyond permission.Prospect security checks are passive and external only. Anything active requires written consent first, and initial scans never touch a system without it.
• It does not hold credentials in chat or in files.Logins and keys move through secure links into secure storage, and per client credentials are never reused across clients.
• It does not confuse its layers.The watchdog is health monitoring and is never sold as a vulnerability scanner. The scanners do the hunting; the watchdog makes sure nothing changes unnoticed between hunts.

A client gets the same two layers: a watchdog that checks the live site around the clock & speaks up only when something changed, & scheduled scanning interpreted against the client's actual stack, so the report that arrives is short, evidenced, and carries a recommended fix. Every fix waits for the client approval, & every action sits in an audit log the client can inspect.

The trade off is the one this case study is named for. A human gate makes remediation slower than an automated tool that patches first & reports later. For a business whose website takes orders & answers customers at 3 in the morning, we judge the checked fix to be worth the wait.

• The Watchdog Is Watching. October 11, 2026. The around-the-clock production check, what it looks at, & the plain statement that it is health monitoring, not a vulnerability scanner.
• Never Hacked: How We Protect You. October 2026. Monitoring scope, severity & owner discipline, release blocking, the approval gate, credential handling, least privilege, & the escalation target.
• Cybersecurity AI Agent build spec. October 11, 2026. Assisted tier, canonical intelligence resources, context matching behavior, evidence backed alerting, & the capability claim rule.
• Cybersecurity AI Agent external audit checklist. October 11, 2026. The six audit areas, the three non negotiables, and the standing two factor authentication rule.
• Case study: How AI Agents Operate at Thanos Labs. Autonomy tiers and the limits every AI agent works under, including read only evidence collection in security work.
• Production watchdog run records. October 2026. The drift catch & its same day correction & clearance.