Tether
A data-flow security governance scheme for AI agent runtimes based on the FalkorDB graph database, which models agent operations with property graphs, performs real-time multi-hop path interception, autonomous rerouting, and verdict memory to achieve data-flow security control under dynamic toolchains.
Core Facts
Timeline (last 90 days)
Tether's end-to-end orchestration demo (6 acts) takes 4.16 seconds in total.
Tether uses a lightweight runtime tag __tether_stamp to track data categories, and uses parameterized Cypher queries to determine whether data can reach an egress of a forbidden trust level within 0 to 4 hops, blocking the call if such a path exists.
Tether's multi-hop data-flow path traversal query latency is under 2 milliseconds.
Tether achieves LLM-free autonomous safe rerouting: when a compliant alternative (such as local_summarizer) in the tool capability graph shares a task type with a violating tool but only writes to an internal sink, the system automatically reroutes.
Tether writes each gating decision into an immutable (:Verdict) node; the agent performs a consultative pre-check before proposing a tool and proactively adopts rerouting candidates when history was blocked, thereby achieving a 0.0% repeat-violation rate.
In Tether's benchmark tests, after revoking a single grant on an upstream tool, active leak paths dropped from 4 to 0 while all 10 legitimate business workflows remained intact.
Tether's benchmark tests show a risky-call interception rate of 6/6 (100%), a legitimate-task false-block rate of 0/10 (0%), and 10/10 legitimate tasks uninterrupted after revocation.
There is no LLM call anywhere in Tether's execution logic; security decisions are made entirely by Python and Cypher queries.
The Tether project is open-sourced on GitHub at the repository sanskar21gusain/tether, and its tech stack includes Python 3.11+, FalkorDB, Docker, FastAPI and Pytest.
The Tether project includes 69 pytest tests and a FastAPI dashboard, but it is still a hackathon product, and its performance in production with respect to policy scale, graph-model maintenance cost and multi-tenant isolation still needs to be verified.
11 more timeline events
All Facts (20)
USDT and USDC account for the vast majority of the stablecoin market share.
65%UnverifiedTether's end-to-end orchestration demo (6 acts) takes 4.16 seconds in total.
50%UnverifiedTether uses a lightweight runtime tag __tether_stamp to track data categories, and uses parameterized Cypher queries to determine whether data can reach an egress of a forbidden trust level within 0 to 4 hops, blocking the call if such a path exists.
50%UnverifiedTether's multi-hop data-flow path traversal query latency is under 2 milliseconds.
50%UnverifiedTether achieves LLM-free autonomous safe rerouting: when a compliant alternative (such as local_summarizer) in the tool capability graph shares a task type with a violating tool but only writes to an internal sink, the system automatically reroutes.
50%UnverifiedTether writes each gating decision into an immutable (:Verdict) node; the agent performs a consultative pre-check before proposing a tool and proactively adopts rerouting candidates when history was blocked, thereby achieving a 0.0% repeat-violation rate.
50%UnverifiedIn Tether's benchmark tests, after revoking a single grant on an upstream tool, active leak paths dropped from 4 to 0 while all 10 legitimate business workflows remained intact.
50%UnverifiedTether's benchmark tests show a risky-call interception rate of 6/6 (100%), a legitimate-task false-block rate of 0/10 (0%), and 10/10 legitimate tasks uninterrupted after revocation.
50%UnverifiedThere is no LLM call anywhere in Tether's execution logic; security decisions are made entirely by Python and Cypher queries.
50%UnverifiedThe Tether project is open-sourced on GitHub at the repository sanskar21gusain/tether, and its tech stack includes Python 3.11+, FalkorDB, Docker, FastAPI and Pytest.
50%UnverifiedThe Tether project includes 69 pytest tests and a FastAPI dashboard, but it is still a hackathon product, and its performance in production with respect to policy scale, graph-model maintenance cost and multi-tenant isolation still needs to be verified.
50%UnverifiedTether holds that the traditional boundary authorization model (checking only individual tool permissions) fails in the face of dynamic toolchains, because individually authorized tools can be combined to form chained data-leak paths.
50%UnverifiedTether uses real physics simulation rather than preset fixed animations, so the ball's movement trajectory is different every time.
50%UnverifiedTether's ball collides with actual application windows on the desktop, bouncing off window edges, rolling along boundaries, and falling into the Dock.
50%UnverifiedTether provides a one-click hide function that can make the ball instantly disappear, to handle situations such as being called on during a meeting.
50%UnverifiedTether requires no account and no subscription; all computation and interaction are done locally, and data never leaves the user's Mac.
50%UnverifiedTether's developer is named Alex.
50%UnverifiedTether is categorized under Mac, productivity, and menu bar application categories.
50%UnverifiedThe Tether team presented the Tether project at the Graph Hacks hackathon; the project uses FalkorDB to model the entire operational universe of an AI Agent as a property graph, intercepting in real time before each tool call to govern data flow.
50%UnverifiedConfirmed speakers at the conference include RJ Scaringe of Rivian, Panos Panay of Amazon, Paolo Ardoino of Tether, Toni Schneider and Rose Wang of Bluesky, Amjad Masad of Replit, and Andrew Feldman of Cerebras.
50%