A stolen key that only works at home is just a string. Tether ties a product's Bedrock access to its own network, then removes the key altogether. This repo is the working demo.
A six-act demo: a startup's product leaks its Bedrock key, a reseller cashes in on it, the key is made useless, and then the key is removed altogether. With a dashboard and a SOC narrator.
Status: built and validated offline (template lint clean, Lambda code parses, the leak exhibits and scan step run locally against a fake key). Not yet deployed or run against a live AWS account. The first deploy is the real test. See "Things the first run will confirm".
This is not a hack against AWS. It is a hack against a company whose product uses an LLM in its backend: a chatbot, a copilot, a support tool, an agent. That backend needs a credential to call the model. If the company stores a long-lived key, the key ends up copied into repos, config files, build logs and laptops. Someone finds a copy and uses it.
Every call the attacker makes is a valid, correctly signed request with the company's own key. AWS does exactly what it is asked, and the company pays. Under the shared responsibility model, where the key lives is the customer's side of the line. That is why the fix is something a partner delivers, not something AWS can switch on for them.
The victim in the demo is LlamaDesk, a made-up startup with a support bot.
| Act | What the audience sees | The point |
|---|---|---|
| 1. The product | A customer question goes in, the bot answers using Bedrock | The victim is an ordinary company with an LLM feature |
| 2. How the key got out | Four exhibits, each with the key sitting in it | Nobody broke in. The key was left lying around |
| 3. The exploit | A reseller serves three of its own paying customers off the key | They want quiet, lasting access to resell. The bill is a side effect |
| 4. Fix 1: network lock | Same reseller, same key: denied. Product still works | A leaked key is useless outside the company's own network |
| 5. Fix 2: no key | The product runs on an IAM role. The exhibits come up empty | The best key is one that does not exist |
| 6. SOC narrator | Plain-language walkthrough of the timeline from the evidence | Detection and a story the customer's team can follow |
| Exhibit | Where the key is | Lesson |
|---|---|---|
| A. Git history | Committed in .env, then deleted in a later commit |
Deleting a secret does not remove it from history |
| B. LLM gateway config | Pasted into litellm_config.yaml to wire up Bedrock |
One file ends up holding every provider's keys |
| C. CI build log | A debug step printed the environment | Logs are read by more people and tools than code is |
| D. The running product's config | Plain-text environment variable | Anyone with read-only account access can see it |
The exhibits are files created locally in CloudShell. Nothing is pushed or published. They contain the
real demo key, so run ./demo.sh park when you are done: it disables the key and deletes them.
Not simulated, but worth saying out loud: poisoned dependencies that read the environment, infected developer laptops, and exposed gateway admin pages.
| Piece | What it is |
|---|---|
| Product (as shipped) | Lambda in a private VPC calling Bedrock through a VPC endpoint with a long-lived IAM user key |
| Product (fixed) | The same code on an IAM role, with no key anywhere |
| Attacker | Lambda outside the VPC acting as a reseller. Our simulated attacker is capped at 100 small calls per run so the demo itself stays cheap. Real attackers have no such limit, and Bedrock has no spending cap by default |
| Network lock | IAM policy that denies Bedrock unless the call arrives through that one VPC endpoint |
| Collector | EventBridge rule on CloudTrail catches every call, labels it, and can apply the lock automatically |
| Dashboard | llmjack-demo: legit, stolen, blocked and role-based calls per minute, who the key was resold to, incidents |
| SOC narrator | Lambda that reads the evidence, writes the walkthrough and answers follow-up questions |
| Cost attribution tags | The product's IAM user and role carry team and project tags, for Bedrock cost allocation by IAM principal. Activate the tags in Billing to see spend per identity |
Everything is named llmjack-demo*. The key and the role can only invoke Amazon Nova models.
- Upload
template.yamlanddemo.shto CloudShell (Actions, Upload file). chmod +x demo.sh && ./demo.sh deploy./demo.sh run- Open the dashboard URL printed by the deploy step.
Pick a Region where the model in ModelId can be called directly. To use an inference profile instead:
./demo.sh deploy ModelId=us.amazon.nova-micro-v1:0 NarratorModelId=us.amazon.nova-lite-v1:0
Self-healing variant, where detection applies the lock on its own: ./demo.sh deploy AutoContain=true
Each act can be run alone: app, exhibits, find, attack, lock, unlock, fixed, narrate.
For extra effect, replay the key from your own laptop: ./demo.sh leak, copy the two export lines
to the laptop, then python3 replay_from_laptop.py.
Ask the narrator anything: ./demo.sh ask "Who was the key resold to and how many calls did each make?"
python3 tether_scan.py (or ./demo.sh scan) is a read-only scan of any AWS account. It answers "how exposed are we
to stolen-key abuse of Bedrock?" with evidence, in four sections that match what AWS asks partners to help with:
| Section | What it checks |
|---|---|
| 1. Access key hygiene | Every active IAM access key: age, last use, never-used and dormant keys, root keys, root MFA, alternate contacts |
| 2. Roles instead of long-lived keys | Long-term Bedrock API keys, IAM users holding broad Bedrock policies, whether a Bedrock VPC endpoint exists to tie access to |
| 3. Detection and cost alerts | GuardDuty AI Protection and Investigation status, Bedrock invocation logging, Cost Anomaly Detection subscribers, Budgets and budget actions |
| 4. Review readiness | Security Hub and unused-access analysis |
It writes tether-scan.md and tether-scan.json. It only calls List, Get and Describe APIs and reads no secret values.
A check it cannot run is reported as UNKNOWN, never as a pass. The report names IAM users, so treat it as confidential.
Status: the key-classification logic is unit-tested, and the credential-report parsing was checked against one live account. The script has not yet been run end to end with a role that can read every service.
| Command | AWS capability | What it does |
|---|---|---|
./demo.sh guardduty-on |
GuardDuty AI Protection (Jul 2026), GuardDuty Investigation preview (Jun 2026) | Enables both features on the detector. Fails in member accounts managed by an organization administrator |
./demo.sh guardduty-samples |
GuardDuty AI Protection | Creates sample cost-harvesting and anomalous-invocation findings. The stack routes real and sample findings to the dashboard and the narrator, labelled |
./demo.sh investigate [finding-id] |
GuardDuty Investigation | Starts an AI investigation of a finding, or of the whole account |
./demo.sh cost-watch |
Cost Anomaly Detection | Creates a Bedrock spend monitor with an immediate alert to the demo SNS topic |
policies/ |
Bedrock API key condition keys (Sep 2025) | A ready policy that blocks long-term Bedrock keys, and the network lock as a reusable policy or SCP |
| IAM tags in the template | Bedrock cost allocation by IAM principal (Apr 2026) | Activate the team and project tags in Billing to see spend per identity |
None of these commands has been run against a live account yet. AI-powered cost investigations (Amazon Q) are a console button on a detected anomaly, so there is nothing to script.
- An attacker running code inside the company's own systems, calling Bedrock through the real endpoint.
- A stolen LLM gateway key. The gateway makes the call from inside the network on the attacker's behalf.
- A stolen key that also has IAM rights and can simply detach the lock.
Act 5 answers the first and third. The gateway case needs gateway hardening: keep it off the public internet, cap spend per key, keep the master key in Secrets Manager.
It is an account or organization setting, so it is left for you to switch on:
GuardDuty console, Protection plans, AI Protection, Enable. Its findings come from learned baselines,
so they will not fire on cue during a ten-minute demo. Use aws guardduty create-sample-findings to show
what a finding looks like, and say so on stage.
- The model IDs are enabled and callable in the chosen Region.
aws:SourceVpcebehaves as expected when the model is called through an inference profile.- CloudTrail delivers denied
Conversecalls to EventBridge, withvpcEndpointIdpresent for in-VPC calls. - The reseller tag in the user agent shows up in CloudTrail (the "resold to" dashboard panel depends on it).
- The Bedrock endpoint is offered in the first Availability Zone (if not, deploy with
AzIndex=1).
./demo.sh parkdisables the key and removes the local exhibits../demo.sh unparkre-enables the key../demo.sh destroydeletes everything.
Running cost is dominated by the one interface VPC endpoint, roughly a cent an hour. Check current pricing.
- Which AWS account should this live in? It needs a role that can create IAM, Lambda, VPC and EventBridge resources.
- Does the network lock also apply to Bedrock's own API keys (bearer tokens)? Unverified. Fallback is an SCP that blocks long-term Bedrock keys.
- Should we add a running LLM gateway (LiteLLM) so the stolen gateway key case can be shown live?
- Which leak routes show up most in real startup cases? Act 2 should lead with those.
- GuardDuty AI Protection: who can enable it, and do we show sample findings or wait for real ones?
- How should this line up with the seller one-pagers for the enablement session?