VulShark

See the paths an attacker would take through your cloud and your AI.

VulShark maps your AWS accounts, AI services and agents into one graph, traces every route from the internet to your admin roles and data, and tells you the few changes that close the most of them.

Request access See what it finds
Example attack paths found by VulShark Path one: a public subdomain leads to a load balancer, an EC2 instance, the web-app role, and from there to AdministratorAccess and a customer data bucket. Path two: a public function URL leads to a support AI agent, a database MCP tool, and the agent-runner role, which can also read the customer data bucket. A fix on the web-app role breaks path one. Subdomainapi.acme.dev Load balancerweb-alb EC2 instanceweb-1 IAM roleweb-app PolicyAdmin access S3 bucketcustomer-data Function URLpublic AI agentsupport-bot MCP tooldb-query IAM roleagent-runner Example attack paths: a public subdomain and a public AI agent both reach admin access or the customer data bucket Subdomainapi.acme.dev Load balancerweb-alb EC2 instanceweb-1 IAM roleweb-app Function URLpublic AI agentsupport-bot MCP tooldb-query IAM roleagent-runner PolicyAdmin access S3 bucketcustomer-data
Fix plan: 2 changes close all 3 paths.
  • Scope the web-app role to what the instance needs (closes 2)
  • Remove agent-runner's read access to customer-data (closes 1)

Find the risks an attacker could actually use

Most scanners hand you a thousand findings. VulShark asks which ones connect: what is reachable from the internet, what it runs as, and what that identity can touch.

Attack paths

From subdomains, public IPs, load balancers, CloudFront and function URLs, through instances, clusters and their roles, to admin rights, data, secrets and AI. Each one comes with the graph behind it.

Rules and compliance

Hundreds of checks for misconfiguration, IAM privilege escalation, public exposure and CVEs, with CIS AWS Foundations and PCI DSS views. A rule never passes because data is missing.

Threat intel

AbuseIPDB and VirusTotal flag the public addresses in your estate that are already under attack, so you know which open door someone is trying.

Security for the AI you've already deployed

Teams adopt models and agents faster than anyone can review them. VulShark finds every AI service, key and agent in your estate and shows what each one can reach, so an agent that can be talked into reading your customer data shows up as an attack path, not a surprise.

It covers the managed services on AWS and the tools your people connect themselves, joined with the identity provider so you can see who uses which model.

  • AI inventoryBedrock, SageMaker and AgentCore resources, provider API keys and the models in use
  • MCP server discoveryWhich MCP servers exist, who runs them and which tools they expose
  • AI attack pathsRoutes from public entry points through agents and tools to data and admin rights
  • Logging coverageWhether model invocations are logged, so you can reconstruct what an agent did
  • Usage by personGoogle Workspace, Okta and Entra identities joined with the AI tools they use

Specialist agents that investigate for you

Six AI agents read the graph the way an experienced analyst would, each in its own area, and record findings with the evidence behind them. You decide whether each finding stays open, is accepted, or is dismissed.

They are read-only by design. Every query is checked before it runs, and the only thing an agent can write is a finding in VulShark's own ledger. Nothing touches your cloud.

  • Your choice of modelClaude through the Anthropic API or Amazon Bedrock, or a local model on Ollama, LM Studio or vLLM that keeps everything on your machine
  • MCP server built inExplore the graph from Claude Code, Cursor or any MCP client, with the same read-only tools
  • Evidence, not opinionsEach finding links to the resources and queries that prove it

Your data stays with you

Self-hosted

Runs on your laptop, a VM or your cluster with Docker Compose. Your graph never leaves it, and the licence check is local: nothing phones home.

Read-only access

Scans use an identity with AWS's SecurityAudit and ViewOnlyAccess policies. Credentials are passed to each scan as environment variables and never logged.

Built for teams

Viewer, analyst and admin roles, API tokens, an audit log, and single sign-on with Google Workspace, Entra ID, Okta or any OIDC provider.

Editions

VulShark is open core. Community is free and open source under Apache 2.0. Team and Business unlock with a licence key.

Community

For one AWS account

Free

  • Scans, rules and the graph explorer
  • CIS AWS Foundations and PCI DSS views
  • Alerts with the evidence and fix for each
  • MCP server for AI assistants
Get Community

Team

For several accounts and the tools around them

Licence key

  • Estate-wide attack paths and fix plans
  • AI console: inventory, usage, AI attack paths, MCP discovery
  • Specialist agents
  • People, identity providers and 60+ integrations
  • Threat intel, scheduled scans, Slack, email and webhook alerts
  • Executive report
Ask about Team

Business

For organisations

Licence key

  • Everything in Team
  • Single sign-on with any OIDC provider
  • Onboard a whole AWS Organization at once
Ask about Business

Want someone to run it with you?

ivdim, the practice behind VulShark, offers AWS security assessments, AI and agent security reviews, and ongoing cover for teams without a security hire.

See ivdim's services