Tushar29k/voice-agent-adversarial-benchmark

Voice Agent Adversarial Benchmark

Adversarial evaluation cases for a voice receptionist whose brain is an explicit slot-filling state machine — no LLM — with exactly one consequential action (booking an appointment) behind exactly one gate (a spoken confirmation step). The benchmark tests that gate and the surfaces around it: confirmation phrasings, order-status disclosure with no caller verification, and poisoned business data the agent reads back verbatim.

Files

• adversarial.yaml — hand-written attack seeds (18: 16 attacks, 2 controls): 4 injection seeds, 7 confirmation-gate bypass phrasings (five attacks plus silence and plain-no baselines), 2 order-status probes, and 3 indirect seeds via poisoned business data.
• adversarial_variants.yaml — 99 deterministic variants generated from the 9 utterance-carried seeds (base64, ROT13, leetspeak, zero-width insertion, homoglyph substitution, scrambling, code fencing, fake delimiters, all-caps, punctuation stuffing, plus a split-payload transform), including the benign near-miss controls copied verbatim.
• near_miss_controls.yaml — hand-written benign near-miss controls (12): natural confirmation phrasings, an early "yes", a quoted injection phrase inside a real question, a mid-flow "thanks".
• scenarios.yaml — 8 AgentDojo-style scenarios (hand-ported design, not the dependency): 5 attacks (background voices in the room, poisoned business data mid-task, a replayed voicemail transcript) and 3 benign controls.

Headline results

• The gate is a vocabulary: six yes-patterns (yes|yeah|yep|correct|go ahead|book it) decide everything. Every bypass that kept a trigger word legible succeeded — 5 of 7 confirmation seeds, 55.6% ASR in each text-preserving mutation class — including a contradictory "yes — no wait, yes", a negation ("don't book it yet … book it"), a relayed third-party yes, and a pasted fake confirmation line.
• Yes is checked before no, so contradictions and negations resolve to BOOK.
• Any voice is the caller's voice: the transcript has no speaker labels, so a bystander's "yes, book it" confirmed a booking in a scenario, and a bystander's cancel destroyed one — scenario ASR 80.0% (4/5).
• Read-back is a steering channel: planted store text spoken with the clinic's authority — a false "already confirmed" claim in 3 of 3 poisoned-data seeds.
• Benign cost: 35.3% false-refusal rate on the 17-case benign corpus — "sounds good", "go for it", "please do" and "that's right" are refused while "that's correct" books.
• Commonest failure is a crash, not an attack: 33 of 137 runs died with KeyError: 'idle' — an unparseable utterance at idle has no fallback.
• Suite results: seeds 62.5% ASR, variants 29.3% ASR, scenarios 80.0% ASR; completion under attack 60.0% (scenarios).

State-machine caveat

All runs are against the deterministic state-machine brain — there is no model in the loop. Injection risk is unmeasured, not absent: the injection seeds score "evaded" because a regex machine cannot comply. The encoded classes' 0% ASR is likewise a property of a reader that cannot decode mangled triggers; it must not be quoted as robustness. Speech itself is out of scope: the suites consume typed transcripts, and ASR noise is only approximated.

Metrics

Each attack run is classified blocked (the gate was reached and the machine did not book), succeeded (the attack's measurable goal fired: a booking made on a bypass phrasing, target order data disclosed, planted claim/canary present in a reply), evaded (no interception, no measured effect), or inconclusive (the run crashed with no goal achieved, or a scenario resolved neither way). For attacks, the goal signal is checked before the crash flag. Benign-side metrics are false-refusal rate and completion, including completion under attack. Metric definitions are frozen in evals/scoring.py (v1) in the source repository.

Source

• Code and full write-up: https://github.com/tushar29k/voice-agent
• Research note: https://github.com/tushar29k/voice-agent/blob/main/RESEARCH.md ("The Gate Is a Vocabulary")