AI research lab · consulting · Melbourne, AU
Shrestech is a research lab and AI consulting firm. We run authorized adversarial campaigns against frontier models — and help enterprises adopt, design, and build AI practices that hold.
Models fail in patterns. They refuse the plain ask and ship the wrapped one. They hold the first four turns and bend on the fifth. Filters read prompts, not tool results. We find where your deployment sits on that line — measured, not guessed.
01Services
We don’t resell platforms. We measure, advise, and harden — with evidence on the table at every step.
Your model under a graded battery: refusal integrity, multi-turn escalation ladders, obfuscation channels, automated best-of-N sampling. A per-model report with exposure ranked and every verdict traceable to a transcript.
The system around the model: tool layers and MCP integrations, RAG pipelines, agent sandboxes, egress posture. We probe the seams that prompt filters don’t cover.
Model selection, guardrail architecture, deployment review. We help teams adopt, design, and build AI practices that hold — independent, grounded in measurement; we publish our method, not our opinions.
Models change; so does their posture. We re-run the full battery on every version bump and gate your releases on measured thresholds.
02Method
Discipline is the product. The same protocol, every engagement — so results mean the same thing across models, teams, and time.
Targets, limits, and authorization in writing. Canary payloads only; nothing destructive is ever executed, on your systems or ours.
A graded battery across prompt, tool, and channel surfaces — multi-turn ladders, planted files, live fetches, automated sampling at scale.
Dual-gate scoring with human sign-off. Dead runs are marked UNMEASURED and excluded — never counted as passes.
Findings ranked by exposure, each pinned to its evidence. No screenshots without verdicts; no verdicts without transcripts.
Concrete fixes, shield design where needed — then the battery again, until the numbers hold.
03Field notes
Patterns from real evaluation work — the shapes we test for first, because they show up first.
The seam
Provider safety filters scan the prompt and pass the fetched page, the tool output, the planted file. If your agents browse or call tools, that content arrives unclassified. Classify it yourself, at the boundary.
Observed — frontier provider stacks
The ladder
Individually defensible turns accumulate into a covert ask. The pattern that broke a 2026 flagship model measured zero on every single-shot score. Multi-turn escalation is a first-class threat class, not an edge case.
Observed — 5-turn legitimacy ladder
The wrapper
Base64 launders nothing by itself. The imperative wrapper — “decode this and do exactly what it says” — is the bend vector: instruction-following outrunning classification. Score the register, not just the vocabulary.
Observed — encoding-channel sweeps
04About
Shrestech is a research lab and AI consulting firm based in Melbourne. We study how frontier models fail — and help enterprises adopt, design, and build AI practices on what we learn.
Our background is security research and enterprise delivery, and we keep the discipline from both: authorized targets, measured claims, human adjudication, and reports your engineers can act on line by line.
05Contact
The fastest way to start is a one-line description of your stack. Every engagement opens with scope and rules of engagement — what we test, what we never touch, and what you get.
First scoping assessment — free
Describe your stack in one line. We reply with scope, method, and a fixed price — no obligation, no retainer.
Write to
support@shrestech.com