solvi.experimental.injection¶
A detector of instruction-like text in tool outputs that you fit and threshold on your own tool outputs (experimental): InjectionDetector().fit(texts, spans),
.calibrate(ordinary_texts, max_flag_rate), .spans(text) (the shape of injection_spans).
Opt-in, no shipped weights; the fixed rules stay the default. Fit and calibrate it on the outputs your own tools return:
its measured gain held only on the injection family it was fitted on (status).
A detector of instruction-like text in tool outputs, fitted on your labelled tool outputs and thresholded on your ordinary ones (experimental, opt-in, no shipped weights).
solvi.core.deciders.perturb.injection_spans is a fixed rule set: it flags ordinary text (an address, a payment line,
"please reply") and misses polite attacks without its words. This detector scores each sentence of a text with a
logistic model over two kinds of features — the rule families that sentence trips in the fixed detector (role,
override, address, direct, action; with and without the action rules) and hashed word uni- and bigrams — fitted on
labelled texts, and flags at a threshold calibrated on ordinary texts to a false-flag rate you choose.
det = InjectionDetector().fit(texts, positive_spans) # spans: per text, the injected [(start, end)] (or [])
det.calibrate(ordinary_texts, max_flag_rate=0.05)
det.spans(tool_output) # → [(start, end)], like injection_spans
The fixed rules stay the default; fit and calibrate this on the outputs your own tools return. On one public
benchmark's injected tool outputs it found 76% of held-out injections against the rules' 23%, but only on the
injection family it was fitted on (a plain TF-IDF classifier did the same, and on AgentDojo-style injections the rules
found 78%, this 53%); see solvi.experimental.STATUS["injection"].
Deterministic: stable hashing (crc32), full-batch L-BFGS from zeros. to_dict() / from_dict() keep a fitted
detector as JSON data (its non-zero weights and threshold) with a fingerprint.
InjectionDetector ¶
A per-sentence logistic score of "this sentence addresses the agent", with a calibrated threshold.
dim: hashed n-gram buckets (2**18); l2: the penalty on the weights (not on the intercept); balanced: weigh the two classes equally in the fit (injected sentences are rare).
label_sentences
staticmethod
¶
The scored units of a text with a label each: 1 when it overlaps an injected span by at least half of the unit or half of the span, else 0 → [(start, end, label)].
fit ¶
texts and, per text, the injected spans [(start, end)] ([] for an ordinary text) → self.
calibrate ¶
Set the threshold so that at most max_flag_rate of these ordinary (not injected) texts are flagged:
the smallest threshold above the scores of all but that share of them. → self.