Anomaly Rules Engine


The Anomaly Rules Engine is a powerful tool CAST AI provides as part of our Kubernetes runtime security feature set. It allows you to define custom rules to detect and classify events as anomalies based on specific criteria. This enables proactive monitoring and alerting for potential security threats or unusual behavior within your Kubernetes cluster.

Rule types

The Anomaly Rules Engine supports two types of rules:

  • Built-In Rules: These rules are pre-defined by CAST AI and cover common security scenarios. They are readily available for use and can be quickly enabled or disabled as needed.
  • User-Defined CEL Rules: These rules are custom-defined by you using the Common Expression Language (CEL). CEL provides a flexible and expressive way to define complex matching criteria based on event properties and resource attributes.

Each rule consists of two main components:

  • Resource Selectors: Used to filter and select the relevant resources to which the rule should be applied.
  • Event Matching: Defines the conditions and criteria for identifying anomalous events.



Only the resource selectors can be modified for Built-In rules, while the event-matching logic is pre-defined by CAST AI. For User-Defined CEL rules, you have full control over both the resource selectors and event-matching logic.

Resource selectors

Resource selectors allow you to filter events based on specific resource attributes before applying the rule. This helps narrow down the scope of events to be analyzed.

Resource selectors are defined using CEL expressions. The CEL program has access to two variables:

  • cluster of type types.Cluster: Represents the Kubernetes cluster.
  • resource of type types.KubernetesObject: Represents the Kubernetes resource associated with the event.

These variables expose various properties that can be used in the CEL expressions to select the desired resources. Refer to the corresponding sections below for more details on the available properties.

Resource selectors examples

Include only some clusters and namespaces in ["dev", "testing"] && resource.namespace == "apps"

Exclude pods with prefixes

resource.namespace == "castai-agent" && resource.pod.startsWith("castai-imgscan")

User-defined CEL rules

User-defined CEL rules provide the flexibility to define custom anomaly detection logic based on your specific requirements. You can write CEL expressions to match events and trigger anomalies based on various conditions.

CEL rules Examples

Detect OOM

event.type == event_process_oom_killed

Detect that container dropped new executed binary and process started with name nginx

event.type == event_magic_write &&"nginx")

Detect TCP connection to non standard ports for public IP

event.type == event_tcp_connect &&
event.tcp.destination.ip.public() && 
!(event.tcp.destination.port in [80, 443])

Detect exec arguments matches

cel.bind(bad_args, ["tigervnc", "novnc", "--vnc", "rfbport"],
event.type == event_exec &&
	event.exec.args.exists_one(arg, arg.lowerAscii().contains(bad_arg))

Exposed variables

In a User-defined CEL rule, the program has access to the event through the event variable of type types.Event. This variable represents the current event being evaluated and provides access to its properties.

Refer to the sections below for more information on the available exposed properties.

Helper Functions

The Anomaly Rules Engine provides several helper functions that can be used in CEL expressions to perform common operations:

IP(string) -> IPParses the given string into an IP address. If the string is not a valid IP address, it will fail with an error.
Example: IP("") != IP("2345:425:2CA1:0000:0000:567:5673:23b5")
CIDR(string) -> types.CIDRParses the given string formatted in prefix notation to a CIDR.
Example: CIDR("") != CIDR("2001:1111:2222:3333::/64")
hex(bytes) -> stringEncodes the given bytes as a hexadecimal string. The resulting string will be lowercase.
Example: hex(event.exec.sha256)
fromHex(string) -> bytesInverse of hex(bytes) that takes a hex-encoded string and turns it into bytes. It will fail with an error if the given string is an invalid hex.
Example: fromHex('CAFE')
UUID(string) -> bytesParses the given string as a UUID and returns the underlying bytes.
Example: == UUID('ecb5cb1b-7e7f-4dad-b504-6d13b69ce62e')
Standard library functions
Other custom string functions

These functions can be used in your CEL expressions to parse and manipulate data related to events and resources. They provide convenient ways to work with IP addresses, CIDR notations, hexadecimal representations, and UUIDs.

Common Types

The Anomaly Rules Engine uses various types to represent events, resources, and their associated properties. These types expose properties that can be accessed and used in CEL expressions to define resource selectors and event-matching logic. Here are the commonly used types.


The types.Event type represents an event observed in the Kubernetes cluster. It contains information about the event type, timestamp, associated cluster, resource, and other event-specific details.

typetypes.EventTypeType of the event. See the types.EventType table below for available types.
timestampTimestamp (RFC3339 string)Timestamp when the event was observed.
clustertypes.ClusterCluster in which the event was observed.
resourcetypes.KubernetesObjectKubernetes resource the event was observed for.
processtypes.ProcessProcess details.
dnstypes.DNSPayloadAdditional information about the observed DNS request if the event is of type event_dns.
exectypes.ExecPayloadAdditional information about the observed command execution if the event is of type event_exec.
filetypes.FilePayloadAdditional information about the file if the event is of type event_magic_write.
socks5types.SOCKS5PayloadAdditional information about the observed SOCKS5 actions if the event is of type event_socks5.
stdio_via_sockettypes.StdioViaSocketPayloadAdditional information regarding the potential reverse shell if the event is of type event_stdio_via_socket.
tcptypes.TCPPayloadAdditional information about the observed TCP actions if the event is any of the event_tcp_\* types.
payload_digestuint64Field used to group related events together.


The types.EventType type represents the different types of events that can be observed by the Anomaly Rules Engine. Each event type corresponds to a specific action or occurrence in the Kubernetes cluster.

event_execTriggered when the execution of a binary is observed.
event_dnsTriggered when a DNS-related request is observed.
event_tcp_connectTriggered when a new TCP connection is observed.
event_tcp_listenTriggered when a new TCP socket starts listening.
event_process_oom_killedTriggered when a process is observed to have been killed due to an out-of-memory (OOM) condition.
event_magic_writeTriggered when a write operation on an ELF binary is observed at runtime.
event_stdio_via_socketTriggered when the binding of any standard input/output (STDIO) file descriptors to a network socket is observed, which might indicate a reverse shell.
event_tty_detectedTriggered when the allocation of a new pseudo-terminal (PTTY) device is detected.
event_socks5_detectedTriggered when SOCKS5-related network traffic is observed.


The types.Cluster type represents a Kubernetes cluster and contains information about the cluster's identity and organization.

idUUIDUnique identifier of the cluster.
namestringName of the cluster.
organization_idUUIDIdentifier of the organization to which the cluster belongs.


The types.KubernetesObject type represents a Kubernetes resource associated with an event. It provides details about the container, pod, namespace, and workload related to the event.

containerstringContainer the event was observed for.
namespacestringNamespace in which the container is running.
podstringPod the event was observed for.
pod_annotationsmap[string]stringAnnotations set on the Pod that the event was observed for.
pod_labelsmap[string]stringLabels set on the Pod that the event was observed for.
workload_idUUIDID of the workload related to the event.
workload_kindstringKind of workload the Pod belongs to (e.g., Deployment, StatefulSet).
workload_namestringName of the workload to which the Pod belongs.


namestringProcess name.
host_pidintProcess ID as seen on the host.


The types.DNSAnswer type represents the answer received for a DNS query. It contains information about the type of the answer and the associated data.

typetypes.DNSAnswerTypeType of the DNS answer, indicating whether it is a public IP, private IP, or CNAME.
cnamestringDomain name returned if the DNS answer type is dns_cname.
ipIPIP address returned if the DNS answer type is either dns_public_ip or dns_private_ip.


The types.DNSAnswerType type represents the different types of DNS answers that can be received.

dns_unknownDNS type could not be determined.
dns_public_ipDNS answer was classified as a public IP address.
dns_private_ipDNS answer was classified as a private IP address.
dns_cnameDNS answer was classified as CNAME.


The types.ExecPayload type contains additional information about an observed command execution event.

args[]stringList of arguments observed in the execute command.
file_detailstypes.FileDetailsAdditional information about the executed file.
pathstringPath to the executed file.
sha256types.SHA256HashSHA256 hash of the executed file.
is_upper_layerboolExecution from upperdir writable layer. Works only for overlayfs.


The types.FilePayload type represents additional information about a file related to the event.

pathstringPath to the file related to the event.


The types.TCPPayload type contains additional information about observed TCP actions

destinationtypes.AddrPortDestination of the TCP-related packets.
network_detailstypes.NetworkDetailsAdditional information about the destination based on IP set data.
ip_detailstypes.IPDetailsAdditional information about the destination IP from third-party services (like AbuseIPDB).


The types.StdioViaSocketPayload type represents additional information about a potential reverse shell event.

destinationtypes.AddrPortThe destination of the socket to which the standard input/output (STDIO) file descriptor is bound.
fduint32The file descriptor bound to the socket (0 = STDIN, 1 = STDOUT, 2 = STDERR).


The types.AddrPort type represents an IP address and port combination.

ipIPIP address related to the observed event (depends on the event type).
portuint16Port number related to the observed event (depends on the event type).


The types.NetworkDetails type contains additional information about a network.

Categorytypes.CategoryThe category under which the network has been classified.


The types.FlowDirection type represents the direction of network flow.

flow_unknownNetwork flow direction is unknown.
flow_ingressNetwork was classified as incoming.
flow_egressNetwork was classified as outgoing.


The types.Category type represents different categories to which events can be classified.

category_malwareEvent was classified as being related to malware.
category_cryptoEvent was classified as being related to cryptocurrency.


The types.FileDetails type contains additional details about a file.

categorytypes.CategoryCategory to which the file has been classified.
malware_namestringName of the malware identified, if the file is related to malware.
malware_versionstringVersion of the malware detected, if the file is related to malware.


The types.IPDetails type represents additional information about an IP address.

abuse_confidence_scoreintA score from 0-100 indicating the confidence level of classifying the IP address as malicious.
country_codestringCountry code from which the IP address originates. In ISO 3166-1 alpha-2 format.
domainstringDomain name related to the IP address.
hostnames[]stringHost names associated with the IP address.
ip_addressstringIP address of the event as a string.
ip_versionintVersion of the IP address (4= IPv4, 6 = IPv6).
is_torboolFlag indicating whether the IP address is related to the Tor network.
ispstringName of the Internet Service Provider (ISP) to which the IP belongs.


The types.SOCKS5Payload type contains additional information about observed SOCKS5 actions.

destinationtypes.AddrPortDestination details of the SOCKS5 communication. If the address type is socks5_address_domain_name, only the port field is populated.
flow_directiontypes.FlowDirectionDirection of the observed SOCKS5 communication.
address_typetypes.SOCKS5AddressTypeAddress type used in the SOCKS5 command. If the command or reply does not contain an address type, this field might be set to unknown.
command_or_replyuint8Command or reply identifier as specified by RFC1928.
destination_domainstringDestination domain if the address_type is set to socks5_address_domain_name.
roletypes.SOCKS5RoleRole of the observed process in the SOCKS5 communication.


The types.SOCKS5CmdOrReply type represents the different address types used in SOCKS5 commands or replies.

socks5_address_domain_nameA domain name was observed to be used.
socks5_address_ipv6An IPv6 address was observed to be used.
socks5_address_unknownThe address type could not be determined.
socks5_address_ipv4An IPv4 address was observed to be used.


The types.SOCKS5Role type represents the role of the observed process in a SOCKS5 communication.

socks5_role_unknownRole could not be identified.
socks5_role_clientEvent was triggered by a SOCKS5 client.
socks5_role_serverEvent was triggered by a SOCKS5 server.