Security
Security, said plainly.
Anchor8 sees every step your agents take and everything they read back, so it handles some of the most sensitive data you have. Here is what we do today, how the product is built to protect that data, and what will be in place before any customer's data reaches us.
SOC 2 is on its way. Until the report exists, no badge goes on this page.
Where your data goes
Your systems
Anchor8, on AWS
Anthropic
Your systems
Anchor8 SDK
Runs inside your code. It sends each step before it runs and each result before your agent reads it, over HTTPS. If Anchor8 can't be reached, it follows the fallback you set, and every fallback is recorded.
Where we stand
We don't hold any customer data yet. This is what's already enforced, what the product is built to do, and what has to be true before the first customer's data comes in.
Enforced today
Tenant isolation, tested on every change.
Every table records which customer a row belongs to, and the database refuses rows from any other. Tests try to cross that line on every change, and fail on any table nobody has checked.
No customer content in our logs.
Logs carry IDs, timings and codes, never prompts, tool arguments or results. A check fails any change that tries to log a banned field.
Secrets and dependencies scanned.
Every change is scanned for leaked secrets, and for known vulnerabilities and licence problems in what we depend on.
Built into the product
A key that's only yours.
Each customer gets its own 256-bit key. Every stored payload is compressed, then encrypted with it (AES-256-GCM), so storage never holds plaintext.
Nobody reads your data by default.
If one of us has to, they ask for access to your account alone, give a written reason, and get one hour. The grant and every record read under it go into your audit log.
A log you can check.
Each entry is chained to the one before it by a hash, so an edited or missing entry shows. Checking the chain needs no key.
Sign-in done properly.
Passwords are hashed with Argon2id, at least 12 characters and checked against common ones. Repeated failures are slowed down without locking anyone out. Sessions end after 12 hours and can be revoked at once.
API keys kept as hashes.
A key is shown once and stored only as a hash. Its a8_ prefix lets secret scanners spot one that leaks.
Before your data arrives
Two-factor sign-in for every account.
Ours included.
Key rotation.
Every encrypted value already records which key version sealed it, so rotating keys won't strand old data.
Deletion on request.
A defined process for deleting a customer's data when they ask, not only when it leaves.
Backups, and a restore we've tested.
A backup nobody has restored from doesn't count.
A security review.
Of sign-in, key handling and encryption, before any of them is called done.
A named hosting region.
We host on Amazon Web Services. The region goes on this page before launch.
Your agents' data
Tool calls and results can carry anything: customer records, credentials, money. This is how we treat them.
- What we receive
- Each step on its way out and each result on its way back. The agent's task, instructions and reasoning (if your agent exposes it) come once per run, and after that only what's new.
- Where it's kept
- In storage on AWS, encrypted with your own key before it's written. Our logs never hold it.
- How long
- For as long as you're a customer, so your audit log stays complete. When you leave, we delete your key, which makes every stored copy unreadable, and then delete the copies.
- What it trains
- Your review decisions train your own model and no one else's. Our base model learns from synthetic and public data only, and no Claude output goes into any training data.
- Who at Anchor8 sees it
- Nobody by default. Access to one customer needs a written reason, lasts an hour and shows up in that customer's audit log. Sales and marketing never get it.
Who else touches it
Two outside companies, and we'll keep this list current.
- Amazon Web Services
- Hosting and storage. Payloads are encrypted with your own key before they're stored.
- Anthropic
- Runs the Claude model behind the Harness, and only when you use it. It receives the context you point the Harness at, and every send is written to your audit log.
Compliance
How we'll meet the privacy laws you work under, and the independent audit that will prove our controls.
- GDPRBefore our first EU customer
- A data processing agreement with every customer, every sub-processor listed on this page, your data kept only while you're a customer and deleted when you ask, and word from us without undue delay if anything goes wrong.
- India's DPDP ActBefore launch
- The same safeguards, held to the Digital Personal Data Protection Act: reasonable security for personal data, prompt notice of a breach, and erasing personal data once its purpose is served.
- SOC 2Planned
- We'll be independently audited: SOC 2 Type I first, then Type II, which tests our controls over months rather than on a single day. The policies and controls on this page are what the auditor will check.
Request our security pack
Our security policy, and answers to your security questionnaire. Tell us who you are and where to send it.
Report a vulnerability
Found a security problem in Anchor8? Email us. A person reads every report.
security@anchor8.netWhat helps us move fast
- What you found, and where
- The steps to reproduce it
- What someone could reach with it
Please give us a chance to fix it before you tell anyone else.
Your agents are already running.
Anchor8 is for one developer, a small team or a whole company. We're letting people in a few at a time. Take a spot in the line and we'll write when it is your turn.

