A deterministic Python gate for LLM record updates

Products·launched

M-Anchor App: local testbed for prompt attacks against a deterministic record-update gate

A local app for testing whether prompt attacks can change a stored record. An LLM proposes record updates, but a deterministic Python gate with no LLM calls decides what actually gets saved, letting users inspect the proposal, decision, and saved state separately. It targets developers testing prompt injection defenses and state integrity guardrails.

A local app for testing whether prompt attacks can change a stored record. An LLM proposes record updates, but a deterministic Python gate with no LLM calls decides what actually gets saved, letting users inspect the proposal, decision, and saved state separately. It targets developers testing prompt injection defenses and state integrity guardrails.

Releases·Tool Release

Bordair Detector: open-source prompt-injection guard for LLM apps with multimodal scanning

Bordair Detector is a free, open-source (Apache-2.0) Python library that protects LLM applications from prompt injections. It uses a two-stage approach: a fast regex layer for obvious cases and a transformer model for ambiguous inputs. It also scans images, PDFs, DOCX, and audio transcripts for hidden injections.

Bordair Detector is a free, open-source (Apache-2.0) Python library that protects LLM applications from prompt injections. It uses a two-stage approach: a fast regex layer for obvious cases and a transformer model for ambiguous inputs. It also scans images, PDFs, DOCX, and audio transcripts for hidden injections.

Products·building

Mastyf Shield: local fail-closed firewall that intercepts AI agent tool calls to block prompt injection attacks

Mastyf Shield is a local security layer that intercepts MCP tool and shell requests from AI agents before execution. Instead of using an LLM to judge safety, it applies deterministic policy rules and fails closed — unpermitted actions send zero bytes to the OS. It targets developers running agents with local system access who face prompt injection risks like AWS credential theft.

Mastyf Shield is a local security layer that intercepts MCP tool and shell requests from AI agents before execution. Instead of using an LLM to judge safety, it applies deterministic policy rules and fails closed — unpermitted actions send zero bytes to the OS. It targets developers running agents with local system access who face prompt injection risks like AWS credential theft.

Products·launched

Doberman: AI watchdog that blocks dangerous LLM commands

Doberman is a two-layer security system that monitors LLM inputs, outputs, and tool calls to prevent dangerous actions like database deletion. It combines deterministic rules with a dynamic layer that learns user preferences, addressing the gap in existing guardrails that miss 11% of threats.

Doberman is a two-layer security system that monitors LLM inputs, outputs, and tool calls to prevent dangerous actions like database deletion. It combines deterministic rules with a dynamic layer that learns user preferences, addressing the gap in existing guardrails that miss 11% of threats.

Releases·Tool Release

Developer releases toolwall, a fail-closed security gateway for AI agent tool calls

A developer built toolwall, a fail-closed checkpoint that sits between an LLM's tool call and its execution, blocking dangerous calls that pass JSON schema validation. Registration acts as an allowlist, and unknown tools, schema or policy violations, budget overruns, and detected secrets are blocked before execution. The developer is inviting the community to stress-test it.

A developer built toolwall, a fail-closed checkpoint that sits between an LLM's tool call and its execution, blocking dangerous calls that pass JSON schema validation. Registration acts as an allowlist, and unknown tools, schema or policy violations, budget overruns, and detected secrets are blocked before execution. The developer is inviting the community to stress-test it.