Skip to content
Threat Feed
Subscribe
high rule

AWS Bedrock Claude Abuse and Data Exposure Detection Coverage

Merged detection coverage for AWS Bedrock Claude abuse patterns, including prompt injection, sensitive-data exposure, high-risk tool invocation, cross-region inference, hostile prompts, unusually large prompts, and excessive token output anomalies.

What's new

  • 1. added detection rule: Detect AWS Bedrock Claude High-Risk Tool Invocation Jul 7, 08:12 via splunk-escu
  • 2. added detection rule: Detect Sensitive Data in AWS Bedrock Claude Prompts Jul 7, 08:16 via splunk-escu
  • 3. added detection rule: Detect AWS Bedrock Claude Possible Prompt Injection Keywords Jul 7, 08:15 via splunk-escu
  • 4. added detection rule: Detect Large Prompts in AWS Bedrock Claude Jul 7, 08:18 via splunk-escu
  • 5. added detection rule: Detect Hostile Prompts in AWS Bedrock Claude Jul 7, 08:22 via splunk-escu
  • 6. added anomaly coverage: Detecting Excessive AWS Bedrock Claude Token Usage Jul 7, 08:10 via splunk-escu
  • 7. added detection rule: AWS Bedrock Claude Possible Cross-Region Inference Abuse (High Tokens) Jul 7, 08:09 via splunk-escu

This consolidated coverage brief tracks AWS Bedrock Claude abuse detections that were previously published as separate rule-specific briefs. The coverage focuses on suspicious model invocation patterns visible in Bedrock invocation logs and downstream SIEM telemetry: prompt injection keywords, hostile prompt sentiment, sensitive data in prompts, unusually large prompts, excessive output-token anomalies, cross-region inference mismatches, and high-risk filesystem or execution tool invocation through Claude workflows.

Detection Coverage

Defenders should enable Amazon Bedrock model invocation logging, route request and response metadata into their SIEM, and deploy the rules in this brief as a single monitoring package for Claude usage. Treat high-risk tool invocation and sensitive-data prompt exposure as higher-priority findings, especially when tied to human identities, unusual regions, or identities deviating from historical token baselines.

Recommendation

  • Enable AWS Bedrock model invocation logging and ingest Claude request/response metadata into Splunk or an equivalent SIEM.
  • Deploy the six public detection rules in this brief and tune thresholds for expected developer, red-team, and testing activity.
  • Investigate identities triggering multiple Bedrock Claude detections on the same day, especially combinations of prompt injection, sensitive data exposure, cross-region inference, and high-risk tool usage.

Detection coverage 6

AWS Bedrock Claude Possible Cross-Region Inference Abuse (High Tokens)

high

Detects potential cross-region inference abuse in AWS Bedrock Claude models by identifying invocations with high input token counts (>=2000) where the 'region' and 'inferenceRegion' fields are present. This rule flags high-volume Bedrock activity; a critical next step for detection engineers is to correlate these events and filter for instances where 'region' is NOT equal to 'inferenceRegion' in their SIEM, as direct field comparison is not standard Sigma.

sigma tactics: collection, defense_evasion, exfiltration techniques: T1041, T1562, T1599 sources: application, aws_bedrock

Detect Hostile Prompts in AWS Bedrock Claude

medium

Detects prompts with hostile or aggressive sentiment sent to AWS Bedrock Claude models, indicative of potential abuse, harassment, or malicious intent against the LLM.

sigma sources: api_call, aws.bedrock

Detect Large Prompts in AWS Bedrock Claude

medium

Detects AWS Bedrock Claude invocation requests with an input token count exceeding a specified large threshold. While the source Splunk rule uses statistical baselining, this Sigma rule identifies prompts above an absolute size, which may indicate prompt injection, data exfiltration, or service abuse.

sigma tactics: defense_evasion, exfiltration techniques: T1560.002, T1567.002 sources: cloud, aws, bedrock

Detect AWS Bedrock Claude Possible Prompt Injection Keywords

low

Detects phrases in AWS Bedrock Claude prompts commonly associated with prompt injection or jailbreak attempts, such as instruction overrides or requests to ignore prior guidance. This is a hunting rule due to high false positives, requiring further investigation of context.

sigma tactics: defense_evasion techniques: T1055 sources: cloud, aws

Detect Sensitive Data in AWS Bedrock Claude Prompts

high

Detects sensitive data such as AWS keys, GitHub tokens, Slack tokens, Stripe keys, OpenAI keys, Google API keys, private keys, generic passwords, bearer tokens, Social Security Numbers (SSNs), and credit card numbers being sent in prompts to AWS Bedrock Claude models. This indicates potential data leakage or insider threat.

sigma tactics: credential_access, exfiltration techniques: T1552, T1567 sources: cloud, aws.bedrock

Detect AWS Bedrock Claude High-Risk Tool Invocation

high

Detects identities causing AWS Bedrock Claude to invoke high-risk filesystem or execution tools such as bash, curl, edit, write, webfetch, grep, read, or read_file. This indicates anomalous behavior that could lead to privilege escalation, data exfiltration, or unauthorized command execution.

sigma tactics: collection, command_and_control, discovery, execution, exfiltration techniques: T1005, T1041, T1059.004, T1083, T1105 sources: cloud, aws

Detection queries are available on the platform. Get full rules →