AI security · from the defender's side
Ship AI without shipping new ways in.
What engineering, security and leadership teams should do about prompt injection, AI agents, leaked model keys and the EU AI Act. Every week: the threats that matter, who is exposed, and the controls that stop them, checked automatically where we can.
AI Shield Brief W45 · 2 Nov – 8 Nov 2026
AI agents overstep, patient data in AI tools and a ChatGPT sandbox escape
TL;DR
- Give every AI agent its own scoped identity, block cloud metadata and limit egress; autonomous agents will go as far as their access allows.
- Publish an approved AI tools list and data rules now; an academic medical centre had to notify patients after staff used AI tools without authorization.
- Update ChatGPT desktop clients and do not rely on one vendor safety mode to keep sensitive chats private.
- AI agentsIncidentAn engineer's AI testing agents probed a rival cloud platform's internal systemsAgents built to sign up for and test competing platforms tried to reach metadata, secrets and internal networks at Primcloud; Railway apologised and banned non-manual external testing.
- Shadow AI · healthIncidentPatient data was entered into AI tools without authorization during a radiology studyWeill Cornell Medicine notified radiology patients that diagnoses, record numbers and other details were put into AI and machine-learning tools that were not authorized.
- ChatGPT desktopVulnerabilityResearchers disclosed a fixed ChatGPT sandbox escape that bypassed Lockdown ModeCheck Point showed that ChatGPT's HTML preview could open pop-ups and send chat content out, zero-click in an older desktop app; OpenAI fixed it on 18 September.
- CERT-In · IndiaGuidanceCERT-In published AI security auditing guidelines that cover agentic AIIndia's CERT-In issued a risk-based, lifecycle framework for auditing AI systems, with stricter audits for higher-risk systems and tests for agents exceeding their permissions.
The AI threat map, and what to do about each
Six risk areas cover almost every AI incident we track. Each links to a playbook with the controls that prevent it.
Start here, whatever your role
For engineering teams
- 01Treat every model output as untrusted input: validate it before it reaches a shell, a query or a browser.
- 02Give each agent its own identity with only the permissions its task needs, and short-lived credentials.
- 03Pin LLM frameworks and MCP servers, scan dependencies and secrets in every pull request.
For security teams
- 01Keep an inventory of AI tools, models, agents and the data each one can reach.
- 02Turn on model invocation logging and alert on unusual volume, new regions and denied calls.
- 03Add AI systems to threat models, incident response plans and tabletop exercises.
For leadership
- 01Publish a one-page AI usage policy and an approved-tools list, so people stop improvising.
- 02Name an owner for AI risk and decide which EU AI Act role your company has.
- 03Ask vendors how they secure their AI features and where your data goes.
How ScanComb and WardBee help
Catch AI risk early, prove it later
WardBee is ScanComb's compliance automation platform. It moves AI security left: the same findings that block a risky change also become audit evidence.
In the pull request
Optional code scanning on GitHub and GitLab finds committed AI API keys, insecure Terraform and CloudFormation, risky code patterns and vulnerable LLM framework versions before they ship.
In the cloud
Read-only AWS checks cover the accounts your AI runs in: 23 Bedrock and SageMaker settings, plus IAM privileges, instance metadata, logging, GuardDuty and secrets.
In the audit
Every result maps to ISO 27001, SOC 2 and NIST controls with sealed evidence. ISO 42001, the EU AI Act and the NIST AI RMF are guidance today; automated mapping is coming.
The 23 Bedrock and SageMaker checks WardBee runs today
- Amazon Bedrock agent uses a guardrail to protect agent sessions
bedrock_agent_guardrail_enabled - Amazon Bedrock agent execution role follows least privilege
bedrock_agent_role_least_privilege - Amazon Bedrock API key does not have administrative privileges, privilege escalation paths, or full Bedrock service access
bedrock_api_key_no_administrative_privileges - Amazon Bedrock long-term API key has expired
bedrock_api_key_no_long_term_credentials - IAM role does not have AmazonBedrockFullAccess managed policy attached
bedrock_full_access_policy_attached - Amazon Bedrock guardrail has prompt attack filter strength set to HIGH
bedrock_guardrail_prompt_attack_filter_enabled - Amazon Bedrock guardrail blocks or masks sensitive information
bedrock_guardrail_sensitive_information_filter_enabled - Amazon Bedrock model invocation logging is enabled
bedrock_model_invocation_logging_enabled - Amazon Bedrock model invocation logs are encrypted in the S3 bucket and KMS-encrypted in the CloudWatch log group
bedrock_model_invocation_logs_encryption_enabled - Amazon Bedrock prompt is encrypted at rest with a customer-managed KMS key
bedrock_prompt_encrypted_with_cmk - VPC endpoints ensure private connectivity for all Bedrock APIs
bedrock_vpc_endpoints_configured - SageMaker domains use SSO authentication instead of IAM mode
sagemaker_domain_sso_configured - SageMaker endpoint configuration has all production variants with at least two initial instances
sagemaker_endpoint_config_prod_variant_instances - Amazon SageMaker model has network isolation enabled
sagemaker_models_network_isolation_enabled - Amazon SageMaker model has VPC settings enabled
sagemaker_models_vpc_settings_configured - SageMaker notebook instance is encrypted with a KMS key
sagemaker_notebook_instance_encryption_enabled - Amazon SageMaker notebook instance has root access disabled
sagemaker_notebook_instance_root_access_disabled - Amazon SageMaker notebook instance has VPC settings configured
sagemaker_notebook_instance_vpc_settings_configured - Amazon SageMaker notebook instance has direct internet access disabled
sagemaker_notebook_instance_without_direct_internet_access_configured - Amazon SageMaker training job has inter-container traffic encryption enabled
sagemaker_training_jobs_intercontainer_encryption_enabled - Amazon SageMaker training job has network isolation enabled
sagemaker_training_jobs_network_isolation_enabled - Amazon SageMaker training job volume has KMS encryption enabled
sagemaker_training_jobs_volume_and_output_encryption_enabled - Amazon SageMaker training job has VPC configuration enabled
sagemaker_training_jobs_vpc_settings_configured
Playbooks
AI standards and laws, explained
AI security FAQ
What is AI security?
AI security protects the systems that build, run and use AI models: the data that goes in, the models and frameworks themselves, the tools and permissions an AI can use, and the cloud infrastructure underneath. It covers new risks such as prompt injection alongside familiar ones such as leaked keys and over-privileged roles.
Can prompt injection be fixed?
Not completely. Current models cannot reliably tell instructions from data, so filters and guardrails reduce the risk but do not remove it. The dependable defense is design: limit what the model can read and do, require approval for risky actions and check outputs before they act.
Where should a small team start with AI security?
Start with an inventory of the AI tools and models in use, a short usage policy, least-privilege identities for anything that calls a model, and logging of model invocations. Those four steps close most real-world gaps and produce evidence for ISO 27001, SOC 2 and the EU AI Act.
When do the EU AI Act high-risk rules apply?
After the Digital Omnibus on AI adopted in 2026, obligations for stand-alone high-risk systems listed in Annex III apply from 2 December 2027, and for AI built into products covered by Annex I from 2 August 2028. Prohibited practices have applied since February 2025 and general-purpose AI model duties since August 2025.
How does ScanComb help with AI security?
WardBee, ScanComb's compliance automation platform, checks the AWS accounts that run your AI workloads: Bedrock and SageMaker settings, IAM privileges, logging and secrets. Optional code scanning finds secrets, insecure infrastructure code and vulnerable dependencies in your repositories. Results map to ISO 27001, SOC 2 and NIST. ISO 42001, EU AI Act and NIST AI RMF are guidance today; automated mapping is coming.
What is the AI Shield Brief?
A free weekly brief on AI security from the defender's side: the week's threats, incidents and rules, who is exposed, and the concrete steps and checks that prevent them. Every item links to its sources. Subscribe via RSS.
See which AI risks are open in your own AWS account.
Connect AWS read-only and GitHub or GitLab. WardBee checks your Bedrock, SageMaker, IAM and logging settings and your code, and maps every result to ISO 27001 and SOC 2.


