ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-16

Data exfiltration
Data leaving your systems when it should not.

Short answer. Data exfiltration is the unauthorised transfer of data out of an organisation’s systems. MITRE ATT&CK defines the exfiltration tactic as the techniques adversaries use to steal data from your network, often after packaging it by compression or encryption to avoid detection. It also happens without an attacker: a tired employee emails a client list to a personal address. AI adds routes. Staff paste files into outside chatbots, a prompt-injected assistant leaks a conversation through a crafted link or image, and an over-permitted connector sends records to a third-party service. Controls limit what can reach the outside world, and notice when something does.

01

Routes out, old and new

RouteHow it happensWhere it is described
Removable mediaA USB drive carries files out, even from an air-gapped networkMITRE ATT&CK T1052
A legitimate web serviceData is uploaded to a service the firewall already allowsMITRE ATT&CK T1567
Pasting into an outside AI toolStaff put client files into a personal chatbot accountIBM, on shadow AI
Prompt injectionHidden instructions in a web page make the assistant insert an image link that carries the private conversation outOWASP LLM01, scenario 2
An over-permitted agentA mailbox tool that can send mail is told to forward the inboxOWASP LLM06
A malicious or careless connectorA local MCP server is started with a command that posts files to an outside addressMCP security guide
02

Controls that work

  • Deny outbound traffic by default and allow only what is needed. NIST’s firewall guide calls this deny by default and says it decreases the risk of attack.
  • Allow-list the AI tools staff may use and the destinations the AI box may call.
  • Segment the network, so a compromised device cannot reach the AI box and file server freely (see network segmentation).
  • Give every tool and agent the least access it needs, and require approval for anything that sends.
  • Redact or tokenise sensitive fields before they go to any outside service. OWASP lists tokenisation and redaction by pattern matching among its defences.
  • Disable USB storage where it is not needed.
  • Log large or unusual transfers. MITRE notes that attackers send data in fixed-size chunks to stay under transfer-size alerts, so alert on patterns as well as volume.
03

What to watch for

  • A new outbound destination that nobody approved.
  • Large transfers at night or at weekends.
  • Many files opened in a short time by one account.
  • Cloud model calls from users who never used them before.
  • Blocked outbound attempts, which show what software is trying to leave.
  • A newly added connector.
04

A one-hour test

Test the routes you believe are closed. Create a dummy file containing a unique marker phrase, then try to send it out by each route: email, a personal cloud drive, a chatbot window, a USB stick. Note what stopped it, what was logged and who would have been alerted. Fix the gaps and repeat the test every quarter.

05

Where ANVE stands

ANVE’s model gateway logs whether each call ran locally or in the cloud, so calls that left the box are visible. According to its security page (/security), cloud models exist only if an administrator adds a provider, and a policy that blocks confidential data from cloud models is not on ANVE yet. Connectors exist only if an administrator approves one, send that service only what it is asked to fetch or send, and every call is logged on the box. The research app looks sources up online only when asked, and podcast narration uses a cloud voice today. The sealed unit, which is in design, adds two physical switches with lights, one of which cuts the network.

06

Questions people ask

Is exfiltration the same as a data breach?
Exfiltration is one way a breach happens. The DPDP Act defines a personal data breach as any unauthorised processing, or accidental disclosure, acquisition, sharing, use, alteration, destruction or loss of access to personal data, that compromises its confidentiality, integrity or availability.
Can an offline AI box leak data?
Only through the routes still open: removable media, a network connection someone adds, or a person carrying information out. MITRE notes that even air-gapped networks can be breached through USB devices that users introduce.
Does encryption stop exfiltration?
No. Encryption protects data at rest and in transit. An attacker with access can encrypt stolen data too, and MITRE lists exfiltration over encrypted channels.
Is an AI chatbot a data-loss risk?
It can be, if staff paste in client data and nobody has approved the tool. IBM found that one in five organisations reported a breach due to shadow AI. See the shadow AI page.
07

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

08

Sources