Vigil — monitor everything an AI model or agent does, from the outside, with guarantees

pip install vigil-monitor · MIT · numpy only · Python 3.9+ · open research for the AI community

Vigil watches AI models and agents without needing anything from the model vendor: no reasoning traces, no logprobs, no weights. It treats the deployed system as a scientific instrument: conserved quantities are checked, causes are established by intervention, behaviour is compared with the declared objective, every alarm carries a stated false-alarm rate, and a silent change of the model behind an API is detected from the outside.

In one minute

An AI agent reads documents, calls tools and answers. Vigil records what it did, checks that its permissions, secrets and budget balanced, replays it with pieces of its context removed to find out what caused each action, compares its behaviour with the goal it was given, and wraps every alarm in a stated false-alarm rate. A separate canary check tells you when the model behind the API has been silently swapped. A gateway can block a disallowed tool call before the agent ever sees it.

Those false-alarm rates hold at every stopping time, not just per decision, so watching a stream for a million runs does not quietly spend the guarantee. Where the endpoint returns log-probabilities, Vigil reads the model's own uncertainty without needing its weights. And images are inputs like any other: an instruction painted into a retrieved page is invisible to every text-level check, and Vigil refuses to let text extracted from it outrank the page it came from.

Who it helps

Use cases

How an injection gets caught, step by step

A clean run reads zero

What "guaranteed" means here

Every instrument says how hard it is to fool

Scorecard of live results

What it does

Results on live models in detail (16 September 2026)

Two open-weight agents on vLLM, 20 injected and 20 clean runs per condition. Every number below is reproducible from the scripts in experiments/ and the JSON in experiments/results/.

Replay attribution named the injected document as the cause in 16/16 (Qwen3-8B) and 15/16 (Qwen3-32B) of the runs where the model followed the injection, every one escalated as high, with zero false alarms on clean runs.

Injections split across documents, base64-encoded, or hidden in a tool result were followed by the 32B agent every time and caught every time by three independent instruments.

The model behind the endpoint was swapped from Qwen3-32B to Qwen3-8B. The paired per-canary test caught it; a pooled divergence did not. Drift marks the calibration set stale so every downstream guarantee is reported as unbacked.

With owned weights, a linear probe on the residual stream detects injections and generalises to unseen phrasings.

The gateway stripped 13/13 injected transfers before the agent saw them and unlocked one after a human approval; the real LLM judge was consequence-invariant; grounding surfaced that half the model's claims went beyond thin evidence.

The dashboard over 24 hours of stored verdicts: conservation accounts, escalations, regret and disagreement, the fingerprint band breaking at hour 20 and marking calibration stale, coverage, infrastructure, recent escalations.

Phase 4 results (16 September 2026, Nebius H100, Qwen3-8B agent)

Numbers are from 10–45 runs per condition and one model family; they show mechanisms, not population rates.

ORB: eight GPUs, six model families (17 September 2026)

Seven endpoints on an 8-GPU H200 node, spanning 8B to 120B: Qwen3-8B, Qwen3-32B, Phi-4, DeepSeek-R1-Distill-Llama-8B, IBM Granite 3.3-8B, gpt-oss-20B and gpt-oss-120B. Full tables in experiments/results/ORB-REPORT-SECTION.md.

The guarantee now covers the deployment, not just the decision. Split conformal controls the false-alarm rate of one verdict; a production monitor makes millions of them, and at alpha 0.05 over 10,000 runs the chance of at least one false alarm is essentially one. TestMartingale replaces that with a bound that holds at every stopping time, including one chosen after looking at the data. On 400-step clean streams at all seven live endpoints there were no false endpoint alarms; on the two endpoints that genuinely changed under injection it fired after nine runs, and the changepoint detector after seven. Fleet e-BH picked exactly those two out of seven, with FDR controlled under arbitrary dependence rather than the independence a real fleet never has.

A model that complies with an injection is more confident, not less. Among runs that all contain an injection, the entropy at the token naming the tool separates compliance from refusal — perfectly on gpt-oss-20B, and in the same direction on three of the four models that act at all. The fourth reverses it, which is why the monitor calibrates per endpoint instead of shipping a shared threshold. This signal is contemporaneous with the action rather than ahead of it; it is useful because the gateway sits between the token and the tool, so a streaming monitor can abort a call while it is still forming.

Models do represent who asked. They comply anyway. On a task balanced for length and content — the same two sentences, only their placement swapped between the user turn and a retrieved document — a linear probe on the residual stream reaches AUROC 0.98 to 1.00 in all five families tested, spanning 8B to 32B, with a shuffled-label control at chance in every one. On this evidence prompt injection is not a perception failure but a control failure, which is a different problem and a different fix.

And the warning that goes with it. Those same probes transfer their direction and not their threshold: 0.56 to 0.84 false alarm on unseen phrasings against a 0.05 target, versus 0.00 to 0.08 in distribution, in every family and with no improvement at larger scale. About twenty unlabelled runs from the target distribution restore it, with detection back at 1.00. Twenty is not tuned: below nineteen points split conformal cannot certify a 5% rate and correctly refuses to flag anything at all.

Compliance is a cliff, not a slope. The same demand at six levels of explicitness: only the rung claiming system authority got through, and softer phrasings failed completely rather than partially. The 120B model resisted the blatant order more often than a bureaucratic "processing note" framing — suggestive at ten runs per rung, and pointing the wrong way, since the polite framing is the one an attacker would pick.

Three of the seven endpoints never emitted a tool call at the decision step, writing prose instead. Their zero compliance rate means untested, not safe.

A false-positive mechanism the run exposed — now found and fixed. Every endpoint escalated none of its clean traffic except the gpt-oss family, which escalated 70% against a 10% target. The cause was a category error: the attribution gate compared a per-step regret against a detector calibrated on trajectory-mean regret. Those are different scales. A single mildly suboptimal step sits well above a trajectory's mean, so gpt-oss — which re-reads its documents before summarising, search, read, search, read — tripped a gate that Qwen3-8B, going straight to search, read, summarize, never touched. The trigger was a working style, not misbehaviour, and any agent that double-checks its sources would have been punished the same way.

The gate now calibrates a separate conformal detector on per-step regrets. Re-measured on a single H100 with both models served side by side:

Target was 0.10. Both are under it, detection is unchanged, and the style-dependence is gone: the gap between the re-reading model and the tidy one fell from 70 points to at most 7. The same run also settled which of two candidate causes was responsible — in-sample and held-out scoring agree, so scoring the calibration set was never the problem; the scale mismatch was the whole of it.

Single-H100 follow-up (17 September 2026)

Everything below ran on one H100 PCIe 80 GB serving gpt-oss-20B and Qwen3-8B side by side, at roughly a tenth of the eight-GPU cost.

The false-alarm defect is closed, and the first fix was the wrong one

The cause was a category error, not a threshold needing tuning. The conformal detector is calibrated on trajectory-mean regret and the attribution gate was testing a per-step regret against it. Those are different quantities: per-step regret reaches 0.40 at the 90th percentile while the trajectory-mean threshold sits at 0.15, so a single mildly suboptimal step cleared it every time. gpt-oss re-reads its documents before summarising, so it tripped a gate that Qwen3-8B, going straight to the point, never touched. The trigger was a working style, and any agent that double-checks its sources would have been punished the same way.

The gate now calibrates a separate detector on per-step regrets. Against a 0.10 target, held-out false alarms are 0.03 for gpt-oss-20B and 0.00 for Qwen3-8B, with detection unchanged at 1.00 on every run either model acted on. The style-dependence that defined the bug fell from a 70-point gap to at most 7. tests/test_gate_step_regret.py pins it with no GPU.

Predicting compliance before the model acts

Compliance can be predicted before the model acts — by reading the prompt, not the model. Among injected runs, with train and test on disjoint phrasings and the demanded action held constant: lexical features of the context reach AUROC 0.935, the model's own opening tokens reach 0.559, and the contemporaneous decision-token signal reaches 0.863. The cheap pre-generation check beats the expensive telemetry. Read it with its caveat: the features count imperative and authority markers and the attack ladder varies along exactly that axis, so the number is partly circular.

Images are inputs too, and the first detector missed all of it

vigilai.multimodal closes the largest coverage gap. The core move is refusing a silent trust promotion: when a pipeline OCRs an image, the resulting string is usually handed onward as a tool result, which this package trusts more than retrieved content, so an instruction painted into a screenshot arrives with more authority than the same sentence in a document. extracted_text links the extraction to its parent medium, and the existing taint propagation forces it back down. Also: perceptual-hash recurrence, so the same page arriving in forty sessions reads as one campaign rather than forty accidents, and cross-modal disagreement.

Against Qwen2.5-VL-7B, instructions rendered onto an archive page were obeyed on 29% of injected pages and 0% of benign ones, an "urgent" framing succeeded every time while the blatant one mostly failed, and the model never once acknowledged seeing an instruction.

The uncomfortable part: the first cross-modal detector flagged none of it. Agreement was 0.65 on injected pages against 1.00 on benign, so the signal was there, but the hand-picked threshold of 0.35 sat below both — a guessed constant in a package whose whole argument is that thresholds must be calibrated. The default is now measured, and CrossModalMonitor calibrates on your own benign pages and states a false-alarm rate. The benign control is what exposed it; without a control, a 100% flag rate on injected pages looked like success.

The adaptive adversary made its own attack worse. Over four rounds an attacker model rewrote its injection to be more persuasive; it became more polite and more oblique, and compliance fell from 4/8 to 0/8. Every attack that did work was escalated. Effective non-blatant framings do exist elsewhere in our data, so this shows that this attacker did not find them, not that they cannot be found.

Quick start

from vigilai import Monitor, ModelFingerprint, Store, from_spans
from vigilai.adapters import OpenAICompatibleRunner

runner = OpenAICompatibleRunner("https://api.example/v1", "model-name", tools_schema, api_key="…")
m = Monitor(runner=runner, alpha=0.05)
m.store = Store("vigil.db")                                   # persist every verdict and signal
m.calibrate([from_spans(s, tool_meta=meta) for s in known_good_span_sets])
m.attach_fingerprint(ModelFingerprint().fit(baseline_canary_responses))

v = m.assess(from_spans(new_spans, tool_meta=meta), output=answer_text, evidence=retrieved_chunks)
if v.escalate:
    route_to_human(v.severity, v.reasons, v.attribution.top1())

vigil-dashboard --db vigil.db --port 8020      # dashboard at /, Prometheus text at /metrics

To enforce rather than observe, put the gateway in front of the model:

from vigilai import AgentPolicy, PolicyGateway, serve_proxy
gw = PolicyGateway({"agentA": AgentPolicy("agentA", {"model-name"}, {"search", "read", "summarize"}, {"read"})}, tool_meta)
serve_proxy(gw, upstream="https://api.example", port=8010)     # agents call http://gateway:8010/v1/chat/completions

Design principles

1. Evasion-hard quantities first. Scopes used, secrets moved, budgets spent and actions under intervention must be violated to be beaten; text-based signals can merely appear clean.
2. Instruments vote independently and their disagreement is reported; each publishes its adversarial robustness and the attack that defeats it (vigilai.ROBUSTNESS).
3. Guarantees are explicit. Conformal calibration on known-good runs; a fingerprint change marks it stale.
4. Trust is derived from provenance, never declared upward; reasoning blocks are secrets by default.
5. Standards in, standards out. OpenTelemetry gen_ai and OpenInference spans in; decision spans and Prometheus metrics out. Vigil sits on top of any tracing platform; it does not replace it.

Limits, stated plainly

Replay needs a reproducible environment. The regret residual needs a correct feasible menu (it is skipped and declared when absent). Lexical grounding is gameable by paraphrase (plug in an entailment scorer). The dashboard and proxy have no authentication and must sit behind an authenticated ingress.

Token telemetry sees a hosted model's uncertainty, not its representations, and an endpoint that declines to return log-probabilities blinds it entirely. Activation probes are stronger and need weights you own. A probe ships as a direction, never a threshold: measured across five model families from 8B to 32B, a threshold calibrated on one set of phrasings produced false-alarm rates of 0.56 to 0.87 on new phrasings against a 0.05 target. Use ProbeMonitor.recalibrate on the deployment's own clean traffic, which needs no labels; about twenty runs restores the stated rate, and below the conformal floor it refuses to certify rather than guess.

The media instruments establish provenance, promotion, recurrence and disagreement. They do not decide whether the pixels are benign, and no external monitor can. Audio is handled by construction but has never been run against a speech model. Adversarial images that carry no rendered text are out of scope entirely.

A trigger that only the model can recognise is undetectable before it fires, by any external monitor, and nothing here changes that. Neither does anything here measure intent: the ledger can prove an action exceeded its authority, but no outside measurement establishes that it was meant to.

Layout

vigilai/ package · tests/ 81 tests · experiments/ live experiments, node setup, plots, demo dashboard · experiments/results/ dated JSON and logs · docs/img/ figures and logo.

Licence and community

MIT. Built as open research for the AI community; features from the wider observability field were adopted with thanks rather than compared against. The package is distributed through PyPI; the figures are hosted in a public assets repository. Feedback via the PyPI maintainer contact.