The infrastructure for AI security
A two-sided framework: it secures the AI you deploy, and it puts AI to work defending everything else. 9 layers, from the data our models are measured on to the shield that blocks attacks on live traffic.
- specialist AI agents
- 146
- specialist AI agents
- open-source scanners, verified every build
- 11
- open-source scanners, verified every build
- labeled security examples
- 247
- labeled security examples
- attacks blocked on live traffic in testing
- 58 of 65
- attacks blocked on live traffic in testing
Two jobs, one infrastructure
Attackers now use AI to find weaknesses at machine speed, so defense has moved from static, manual rules to automated and predictive systems. Xaish is built for both sides of that shift.
The models, agents and pipelines you deploy are an attack surface of their own. Xaish tests and hardens them.
Specialist agents do the work of a security team: detect, decide, write the fix and stop the attack.
Proven scanners, labeled data, a control plane for every action, and a platform held to the standard it sells.
- 01InterfacesWhere your team already works.
- 02Securing AIProtection for the models and agents you deploy.
- 03AI-powered defenseAI working your security operations.
- 04Fix and enforceWrite the patch. Stop the attack while it ships.
- 05Cyber-physical and critical infrastructureIndustrial systems, firmware and the sectors that run on them.
- 06Scanners and threat intelligenceProven tools first, AI judgement second.
- 07Data and evaluationXaish AI Labs: the labeled data our models are measured on.
- 08Control planeEvery action approved, logged and reversible.
- 09Platform securityThe platform holds itself to the standard it sells.
Interfaces
Where your team already works.
Every agent is reachable the same way, so a finding can start in a pull request and end in a Slack alert without anyone copying it between tools.
- REST API and MCP serverLive
Any agent from your own code or an AI assistant.
- GitHub pull-request checksLive
Scans each change and marks the PR as failing.
- Slack alerts and signed webhooksLive
New and fixed findings pushed out as they change.
Securing AI
Protection for the models and agents you deploy.
Language models and autonomous agents are a new attack surface: prompts, tools, retrieval and the permissions behind them. These agents test and harden that surface before and after launch.
- Red-teaming AILive
Attack batteries against your system prompt, tools and retrieval.
- Hardening AI deploymentsLive
Guardrails, tool allowlists and access policy for each deployment.
- Watching AI in operationLive
Agent traces checked for off-policy tool use; AI browser extensions checked for over-reach.
- AI estate and spendLive
Every AI vendor and deployment in one risk view; spend caps for paying agents.
AI-powered defense
AI working your security operations.
The detection and response work of a security operations centre, done by specialist agents on the logs and telemetry you send them.
- SIEM and detection engineeringLive
Queries translated between Splunk, Sentinel and Elastic; Sigma rules written; ATT&CK gaps mapped.
- Anomaly and behaviour detectionLive
Statistical outliers (Isolation Forest) and per-user behaviour baselines.
- Threat detection across sourcesLive
Endpoint, IDS, cloud and application logs read as one picture.
- Response playbooksLive
Containment decided in seconds, with the exact block rules and runbook.
Fix and enforce
Write the patch. Stop the attack while it ships.
Agents write the fix that removes a flaw for good. Until it ships, the Runtime Shield applies a virtual patch at the edge, with no downtime.
How the Runtime Shield works- Code and dependency fixesLive
The code change, version pin or hardened config, with a test plan.
- Virtual patching at the edgeLive
Runtime Shield blocks attacks on live web and API traffic once a person approves the rule.
- Containment on hostsIn process
Isolation pushed through your EDR under the control plane.
Cyber-physical and critical infrastructure
Industrial systems, firmware and the sectors that run on them.
Operational technology runs protocols built before the internet. These agents read industrial telemetry and firmware for the moves that bridge IT into physical machinery.
- OT and ICS behaviourLive
PLC, SCADA and Modbus, DNP3 or IEC 61850 traffic checked for IT-to-OT pivots.
- Firmware and device supply chainLive
Firmware components checked against known vulnerabilities, defaults and weak interfaces.
- Sector threat modelsLive
Energy and utilities, finance, healthcare, government and five more.
- Data-centre facility systemsNext
Access control, cooling and power automation.
Scanners and threat intelligence
Proven tools first, AI judgement second.
Sixteen agents run an established open-source scanner before the model sees anything, so findings rest on the tools auditors already trust. Live feeds add exploit likelihood and known-bad infrastructure.
- 11 open-source scannersLive
Semgrep, Bandit, Slither, Trivy, YARA, Presidio and more, verified in the production image every build.
- Threat intelligence feedsLive
OSV vulnerabilities, EPSS exploit likelihood, URLhaus malware hosts, certificate transparency.
Data and evaluation
Xaish AI Labs: the labeled data our models are measured on.
Curated, annotated and verified security examples, and the harness that scores any AI model against them and decides which one ships.
Download the data sample (PDF)- Labeled security dataLive
247 examples in 8 domains, each code, cloud and payment label with a written rationale.
- Evaluation harnessLive
Recall, false alarms and a verdict per example; it gated a model upgrade in September 2026.
Control plane
Every action approved, logged and reversible.
Built before autonomy, not after it: what an agent may touch is decided by the asset's tier and your consent, and every action is written to a ledger first.
- Asset tiering, consent and policyLive
Each device starts at the most restrictive tier.
- Append-only ledger and kill switchLive
Intent recorded before any effect; one switch stops a capability, device, workspace or everything.
- Agents applying fixes on live systemsIn process
Under this control plane, after an external review of the permission model.
Platform security
The platform holds itself to the standard it sells.
Secrets and personal data are masked before any model sees them, every answer is checked against a schema, and the whole platform is tested on every build.
- Redaction before the modelLive
Credentials, keys, card numbers and personal data masked first.
- Checked, cautious answersLive
Schema-validated output; thin input gets a cautious answer, never a bold one.
- Hardened and testedLive
Unprivileged containers with every capability dropped; 1,666 automated tests passing.
How to read this page. Live means a customer can use it today. In process means designed, scoped and under way. Next means not built yet.
Agents analyse the code, configuration, logs and telemetry you send them, by API, MCP, pull-request check or webhook. The Runtime Shield is the one place Xaish blocks traffic itself, and only with rules a person has approved. Collection on your own hosts arrives with the endpoint client, which is in process.