Your team keeps using ChatGPT, Claude, Copilot and Gemini. Argos swaps the sensitive values for safe placeholders on the way out, and puts the real ones back on the way in. The model does its work. Your data stays yours.
One roundtrip: tokenize, shield, forward, rehydrate.
The proxy rewrites sensitive bytes before they leave the box and restores them in the response, so your users never notice. Here is the whole trip.
device
any app, any process
SSN 545-12-8443
Argos proxy
detect → tokenize → shield → forward
[SSN:1037]
LLM
OpenAI · Anthropic · …
The raw value crosses the first hop only. Past the proxy, the model works with placeholders.
01 · Tokenize
Every detected span (names, IDs, keys, medical codes) is swapped for a stable placeholder before anything leaves the device.
02 · Shield
Sensitive figures (revenue, headcount and other business numbers) are protected before anything leaves the device. The model still does useful work; the real figures stay home.
03 · Forward
The tokenized payload travels to the model over the normal wire. To the AI it is an ordinary, fully readable prompt.
04 · Rehydrate
The response comes back and placeholders are restored to the real values. Your team never notices the trip.
Regex catches the obvious. K-anonymity catches the rest.
Detection runs end-to-end in under 50ms. Nothing is sent off-device for analysis: every layer ships with the agent.
Layer 01
Pattern matching
SSN, credit card, IBAN, ICD-10, API keys, JWT, private keys, DB connection strings, emails. Deterministic, sub-millisecond.
Layer 02
Semantic detection
AI-based detection for personal names, medical conditions, minors and special categories that escape simple pattern matching.
Layer 03 · K-anonymity
The combination attack. Where pattern matching stops, re-identification analysis begins.
"John Smith" is safe. "94304" is safe. "1987-03-12" is safe. Put them in one prompt and 87% of the US population is uniquely identifiable. Argos models re-identification risk using k-anonymity and HIPAA Safe Harbor principles, and blocks the dangerous combinations that pattern matching alone misses.
# prompt: "patient John Smith, DOB 1987-03-12, ZIP 94304, ICD F32.1" findings: 4 · re-identification risk: HIGH combination: name + DOB + ZIP + diagnosis verdict: CRITICAL → action: BLOCK
Always
Metadata only
The central reporter sends counts, scores and labels, never raw findings. The PII string never crosses the network.
Both ways
Response-side scan
Models hallucinate emails and paraphrase names. Argos independently rescans the model's output and logs it as a separate model-generated PII event.
0
LLM endpoints monitored
OpenAI, Anthropic, Gemini, Mistral, Cohere, Groq and 10 more.
0
Detection layers
Pattern matching, then semantic detection, then k-anonymity.
0ms
P50 detection latency
Optimized pipeline engineered for imperceptible overhead.
Every installation includes Sentinel. It aggregates anonymized metadata from every reporting agent: org-wide visibility with zero raw data exposure. Seven views, from fleet overview to board-ready compliance reports.
01
Org Overview
Total scans, devices, blocked, redacted, risk distribution, 7-day trend.
02
Device Browser
Per-device peak risk, blocked count, last seen, sortable across every agent.
03
Departments
Engineering vs. Sales vs. Data Science: where the leaks live.
04
Org-wide Scan Log
Every intercepted request with domain, model, regulations, action.