Three findings from one claims portal
ClaimsPortal is a test system we built to behave like the legacy apps we meet on client work: a policyholder claims portal with planted problems. These are three of the things the skills found in it, and in the AWS account around it.
A2Secrets and credential exposure
A key the working tree no longer tells the truth about
ClaimsPortal keeps an AWS key pair in an env file that is committed. A later commit deletes another secrets file. The key in that deleted file is still in git history, and a rejected bearer token is written to the log.
Found. The scan reports the committed key, the history that still holds a deleted secret, the token in the logs, and the absence of a scanner that would stop the next one.
Fix. Rotate the key with the team that uses it, plan the history clean-up, and add a scan that runs before a commit is accepted.
claimsportal-node/.env AWS_ACCESS_KEY_ID=AKIA************MPLEA bill that does not match the machines you think you run
The cost audit reads the AWS account, not the application repository. On the sample account the audit ships with, three lines are enough to see the shape of the bill.
Found. Idle address, oversized compute, and storage left on the expensive tier. The same pass also flags NAT traffic a VPC endpoint would carry, cross-AZ transfer, and On-Demand compute with no Savings Plan.
Fix. Release the idle address, step the instance down one size and watch it, and give the log bucket a lifecycle.
Sample account, monthly rand Stopped instance, address still attached R1 850 / month Web tier m5.2xlarge, 12% average CPU R4 200 / month Log archive with no colder tier R9 600 / monthF3LLM integration security and cost
The claimant writes the prompt, and the model names the payout
Filing a claim asks a model to summarise it. The claimant’s own words are appended to the prompt, with their ID number and medical notes. The reply is stored and shown as HTML. If the model suggests a payout, that number is used.
Found. Untrusted text in the prompt, model HTML rendered on the claim page, a model-chosen amount in the payment, no token cap, and no timeout if the provider is down. Submission waits on that call.
Fix. Separate the claimant’s text from the instructions, render the summary as text, and keep the payout calculation on rules the model cannot set.
server/llm/prompt.js and server/claims/payout.js Description from claimant: ${claim.description} payout = Math.min(payout, summary.recommendedPayout)
Run the same skills on your own system
Sign in with your email and download two skills free while we take feedback from the first teams.