The prompt had other readers
Dragos seeded two million AI prompts with canary credentials and watched who used them. At least three inference providers had been breached for months, resellers were logging traffic, and privacy flags didn't stop any of it. OpenRouter dropped three providers on November 22. A coding agent sends whatever it reads to its model. In Bromure Agentic Coding, what it reads is stand-ins, so a stolen prompt has nothing in it to use.
Between the agent and the model, a prompt passes through routers, front doors, inference servers and log stores. Dragos found thieves at every one of them. Whatever the agent had read went with the prompt.
Picture a Tuesday afternoon debugging session. A test fails with an
AWS error, and the agent runs cat .env to check the region. The file
holds the access key, the secret, a database password and a Hugging
Face token. All of it goes into the next request to the model, because
that is how an agent works. The request goes to whichever provider the
router picked that minute, maybe one serving an open-weight model at
half the price. The developer gets their answer and goes back to work.
According to research Dragos published this week, a request like that could have been copied at the provider's front door, read off a compromised inference server, scraped by a model lab days later or logged by a reseller running a cheap endpoint from a home connection. In each case the attacker had to do nothing more than read it.
Two million prompts with tripwires in them
On November 19, Dragos chief research officer HD Moore published the results of a five-month experiment. From July to November, the team sent about two million prompts to popular AI services. Each prompt carried unique canaries, also called tripwires: fake data that nobody has a legitimate reason to use, so any use of one is proof that someone read the prompt. They planted web links, DNS names, git repository URLs, SSH passwords, and real AWS keys tied to empty, fully logged accounts. Every canary traced back to the exact prompt, model, router and privacy setting that carried it.
What came back:
- Three hosting providers were breached, and none of them knew. The earliest compromise began in mid-July. In the largest, the attacker took a couple hundred AWS keys, each sent in a single request during short windows in late October. Some came from requests the provider rejected before any model ran, so the copying happened at the front door, before any check.
- The keys were used fast. Some were tested within hours. One key seeded on October 21 was used the next day and was still in use three weeks later. The attacker listed S3 buckets, searched Secrets Manager and Parameter Store, created Bedrock API keys and added IAM users to keep access, for tens of thousands of AWS calls in total. On November 1 they brought in ARTEX, an AI-driven penetration testing framework, to check which keys still worked.
- Resellers logged what passed through. On one inference marketplace, sellers listed models at the lowest price, turned on request logging and waited. One person tried SSH root logins by hand with passwords taken from prompts, then tested AWS keys from the same machine. One router forwarded part of its traffic to that marketplace without telling its customers.
- Privacy flags didn't hold. A Chinese AI lab's systems fetched the secret links from more than half of the prompts sent to its models, including about two-thirds of requests marked "no data collection" or zero data retention (ZDR). A git client tried to clone private repositories mentioned in those prompts. One large U.S. provider was still crawling a link two months after a single request.
Moore's summary of the agent problem fits in one line: "AI agents will happily send whatever they can read, including credentials."
The provider reports
The providers have started to confirm it. Phala's preliminary report says an attacker took control of an inference VM serving DeepSeek V4 Flash 0731 and could see request data while it was processed. Its advice: treat every request sent to that model between October 21 and November 8 as exposed, and "rotate any API keys, tokens, passwords, or other secrets included in those requests." Canary tests confirmed misuse of credentials sent on October 21 and 22.
Together AI's notice, published November 21, describes a different route in. Prompt samples from June and July requests to MiniMax M3, GLM-5.2 and Kimi K2.6 sat in a private Hugging Face repository used for testing. Then "an AI coding agent committed a credential that provided access to the private Hugging Face repository" to a public GitHub repository. Together's scanner caught the commit and reverted it, but someone used the credential during the exposure window, downloaded the samples, searched them for credentials to other services and tested those. Five of the credentials they tried were Dragos canaries.
On November 22, OpenRouter removed three providers and took 125 endpoints off its ZDR list, pending validation. It says the findings showed no breach of OpenRouter itself. The Hacker News has the full roundup.
Two failures repeat across these reports. A secret reached the prompt because the agent could read it. And in Together's case, a secret reached a public repository because the agent could read it and push it.
Why a privacy flag can't fix this
Dragos's recommendations are sensible: self-host if you can, go direct to a first-tier provider, pick your router's upstreams yourself, seed canaries, turn on secret filtering. Each one shrinks the list of readers. The prompt still holds whatever the agent read. Moore is blunt about ZDR: "A privacy flag doesn't help when the attacker is already inside."
Secret filtering is a pattern match on the way out. It catches
AKIA… and ghp_… and misses a password with no prefix, the kind the
marketplace reader tried against SSH. A coding
agent is built to read files, environment variables and command output,
and to put what it read into the next request. To change what the
prompt holds, you have to change what the agent can read.
The same session in Bromure Agentic Coding
Bromure Agentic Coding runs Claude Code, Codex, Grok, Kimi and omp in Ubuntu VMs on your Mac, one VM per machine. Every HTTPS request the VM makes passes through a proxy on the Mac, and that proxy is the only place your real credentials exist.
The .env the agent reads holds stand-ins
Import env file… in the machine's Credentials pane reads your
project's .env. It maps names it recognizes, such as
AWS_ACCESS_KEY_ID, GH_TOKEN or ANTHROPIC_API_KEY, to the right
credential type, and turns the rest, like that HF_TOKEN, into
Other API keys scoped to the host you name, here huggingface.co.
Each is exported in the VM under its usual variable name, holding a
fake shaped like the real thing: ghp_ and 36 characters for GitHub,
brm_… for a generic token, brm-db-… for a database secret. The
proxy swaps the real value in only on requests headed to the
credential's own host.
So when the agent runs cat .env or printenv, the conversation it
sends to the model carries fakes. Every reader on Dragos's map gets a
copy of the prompt and nothing they can use. A fake replayed from an
attacker's server never passes through your Mac's proxy, so it reaches
the real API as a fake and fails authentication.
The AWS key in the prompt can't sign anything
The front-door thief took AWS keys, and AWS shows clearly how the
boundary works. In a Bromure VM, ~/.aws/config points at a helper
that hands out your real access key ID paired with a fake secret.
Every SDK and the aws CLI sign requests with that fake secret. On the
Mac, the proxy strips the signature and re-signs with the real secret
before the request leaves. Lift the pair from a prompt and the only
secret you get is the fake. AWS answers every call signed with it with
InvalidSignatureException. The S3 listing, the Secrets Manager
search, the new Bedrock keys and IAM users in Dragos's logs all needed
that secret.
SSH works the same way. Private key bytes never enter the VM. The guest can ask the Mac for a signature, and it can't read the key, so no transcript ever contains one.
Personal data gets stand-ins too
Dragos's canaries were credentials and links, but the prompts they
rode in also carry whatever the agent read from a customer table or a
support export. PII Protection handles that part. Before a request
leaves for the model provider, the proxy finds names, email addresses,
phone numbers, card numbers, government IDs and street lines, using an
on-device model, and swaps each one for a realistic stand-in. The reply
gets the real values back before the agent sees it. The provider, and
anyone reading its traffic, sees @example.com addresses and test card
numbers. Each swap is counted in the Security Timeline with the
destination host, and the values themselves are never logged.
The commit holds a fake, and the push waits
Together's leak took two steps: a coding agent had a real Hugging Face credential, and it pushed that credential to a public repository. In a Bromure machine, neither step lands.
The token the agent can see is the brm_… stand-in, so a commit
carries the stand-in. Anyone who scrapes it from GitHub has a string
that only turns into a real token inside your Mac's proxy, on a request
to huggingface.co from your VM. From their own machine, Hugging Face
rejects it.
Before the commit goes anywhere, the push itself stops at the
Guardrails write policy. New machines default to Prompt before
write for GitHub, GitLab and Bitbucket: reads pass, and each write,
git push included, pauses for a dialog on the Mac that names the
operation. The policy runs in the proxy, outside the VM, so an agent
can't switch it off. An unexpected push to a public repository becomes
a question you answer before the commit is published.
If someone did paste a real token into a file in the VM, the proxy
flags unswapped, real-looking credentials in outbound headers: values
with known prefixes such as sk-ant-, ghp_ or AKIA, and long opaque
tokens. They show up under Leaks only in the Trace Inspector and in
bromure-cli trace leaks. Strict credential mode, also in
Guardrails, blocks any credential Bromure didn't issue itself.
You choose who reads the prompt
Dragos's first recommendation is to self-host. Its price list starts with a Mac Studio with 512 GB of memory. Bromure's Models board has an On-device provider that runs open-weight coding models on the Mac's GPU with Apple's MLX, with nothing to install, and every agent can pick one. A turn served on-device never leaves the Mac. A Custom server entry points agents at any OpenAI-compatible server on your network, such as vLLM, Ollama or LM Studio, so a team can share one GPU box it controls.
When you do use a cloud model, the Models board is also where you pick the route. You can point each agent at a first-tier provider, with its own sign-in or API key, which is Dragos's second recommendation. That key never enters the VM either. A router like OpenRouter is there when you want it, as a choice you make per agent.
When a provider says "rotate what you sent"
Phala asks customers to rotate every secret included in requests to one model between October 21 and November 8. Most teams can't answer which secrets those were, because nobody kept a record of what their agents sent.
Bromure keeps that record on your Mac. At the default AI request
details level, the proxy stores every exchange with a known model
host, encrypted under a key in your login Keychain. The Trace Inspector
filters by host and by Conversations only, and
bromure-cli trace hostnames lists every host each machine contacted.
You can open the requests from the exposure window and read what went
out, request by request. In a Bromure machine, what went out was stand-ins.
The agent reads fakes
Imported .env values, config files and environment variables hold stand-ins. Real secrets are added on the Mac, only toward their own host.
Stolen prompts open nothing
A fake from a prompt fails upstream. A lifted AWS key ID comes with a fake secret, so every signature is rejected.
Pushes wait for you
Prompt before write pauses every git push for approval, enforced in the proxy. A committed token is a stand-in anyway.
You pick the reader
On-device models keep the prompt on the Mac. Cloud routes are a per-agent choice, and the Mac keeps a trace of every request.
What to do this week
- If you sent requests to DeepSeek V4 Flash 0731 through Phala between October 21 and November 8, follow Phala's guidance and rotate any secret those requests could have carried.
- If you used Together AI's serverless MiniMax M3, GLM-5.2 or Kimi K2.6 in June or July, read Together's notice and turn on ZDR in your account settings.
- If you route through OpenRouter and rely on ZDR, check whether your endpoints are among the 125 it removed.
- Assume that any secret an agent read went to every hop between it
and the model. Rotate long-lived keys that have sat in a
.envan agent could open. - Add
artex-agentandartex-judge-r121to your CloudTrail detections, as Dragos suggests.
Then move the secrets out of the agent's reach for good. Import the
.env into a Bromure machine, leave GitHub on Prompt before write,
and point the agents at the provider you chose. The wire boundary, AWS
re-signing and .env import are covered in
Credentials & the Wire Boundary,
personal data in PII Protection,
on-device and custom servers in
Models & Providers, and the local
record in Tracing & Audit.