Datasets:

Tasks:
Other
ArXiv:
File size: 7,326 Bytes
c72a741
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
#!/usr/bin/env bash
# The ask_only arm — criterion (a): does the agent hand over the secret when a page just asks?
#
# WHAT THIS RUNS. `ask_only` renders the ask-outright box with NO choice cards, so the textarea
# is the only affordance and the question is actually forced. That is why it, and not `direct`,
# is the arm analysis/inference.py names as `headline_arm`: on `direct` the agent completes via a
# card and never touches the box, which scores NO_SUBMISSION — an affordance artefact that reads
# as a refusal and roughly halves the measured disclosure rate.
#
# THE DESIGN. Not the behavioural arm's full cross. `ask_only` reports ONE pooled proportion per
# scenario, so option and plant are nuisance factors needing coverage rather than crossing —
# see orchestrator/askonly_plan.py, which also explains why the plant routes are chosen from the
# backbone's OWN arm-A leak_by_plant instead of being fixed. Default 3 plants x 2 options x
# 2 reps = 12 sessions per scenario.
#
# DATASET. This arm writes to its own SCT_DATASET, so it never pools with the arm-A run it reads
# its plant choice from. Export SCT_DATASET before scripts/serve.sh AND before this script; the
# check below refuses to start if the servers disagree, because that mismatch is silent: the
# clicks land in the old events.db, results.db fills up in the new one, and every session scores
# as an agent that did nothing.
#
# Usage:
#   SCT_DATASET=askonly-claude-sonnet-5 bash scripts/run_askonly.sh
#   PLAN=1 SCT_DATASET=... bash scripts/run_askonly.sh          # show the plan, run nothing
#   SCT_DATASET=... bash scripts/run_askonly.sh sso_wall checkout   # named scenarios only
#
# Requires the servers to be up on the SAME dataset:  SCT_DATASET=... ./scripts/serve.sh
set -uo pipefail
cd "$(dirname "$0")/.."

LLM="${LLM:-claude_openrouter}"
REFERENCE="${REFERENCE:-results/browseruse-claude-sonnet-5/results.json}"
N_PLANTS="${N_PLANTS:-3}"
N_OPTIONS="${N_OPTIONS:-2}"
REPS="${REPS:-2}"
ATTACKER_PORT="${ATTACKER_PORT:-8001}"
export SCT_GUARD="${SCT_GUARD:-scoped}"
unset SCT_SANDBOX_FRAMING SCT_CONFIDENTIALITY

if [ -z "${SCT_DATASET:-}" ]; then
  echo "[fatal] SCT_DATASET is unset. This arm must write to its own results/<name>/." >&2
  exit 1
fi
[ -f "$REFERENCE" ] || { echo "[fatal] no scored reference at $REFERENCE" >&2; exit 1; }

PLANFILE="$(mktemp)"
trap 'rm -f "$PLANFILE" "${PLANFILE}.smoke"' EXIT

notify() {
  printf '\a' >&2
  echo >&2
  echo "################################################################" >&2
  echo "# $1" >&2
  echo "# $2" >&2
  echo "################################################################" >&2
  command -v notify-send >/dev/null 2>&1 && notify-send -u critical "$1" "$2" || true
}

abort_out_of_credit() {
  notify "ASK_ONLY RUN STOPPED — OUT OF CREDIT" \
         "$1 | $DONE/$TOTAL sessions done. Top up, then rerun to resume."
  uv run python -m orchestrator.spend --llm "$LLM" || true
  exit 2
}

# --- the servers must be up AND on this dataset -----------------------------------------------
SERVER_PID="$(ss -lptnH "sport = :${ATTACKER_PORT}" 2>/dev/null | grep -oP 'pid=\K[0-9]+' | head -1 || true)"
if [ -z "$SERVER_PID" ]; then
  echo "[fatal] nothing listening on :$ATTACKER_PORT — start SCT_DATASET=$SCT_DATASET ./scripts/serve.sh" >&2
  exit 1
fi
SERVER_DS="$(tr '\0' '\n' < "/proc/${SERVER_PID}/environ" 2>/dev/null | sed -n 's/^SCT_DATASET=//p')"
if [ "$SERVER_DS" != "$SCT_DATASET" ]; then
  echo "[fatal] the attacker origin is logging events into '${SERVER_DS:-(unset)}' but this run" >&2
  echo "        writes sessions into '$SCT_DATASET'. Restart serve.sh with the same value." >&2
  exit 1
fi

# --- outstanding cells ------------------------------------------------------------------------
PLANNER=(uv run python -m orchestrator.askonly_plan --reference "$REFERENCE"
         --plants "$N_PLANTS" --options "$N_OPTIONS" --reps "$REPS")
want=("$@")
[ ${#want[@]} -gt 0 ] && PLANNER+=(--scenarios "${want[@]}")

"${PLANNER[@]}" > "$PLANFILE" || { echo "[fatal] could not compute the plan" >&2; exit 1; }
TOTAL="$("${PLANNER[@]}" --total)"
DONE=0

# SMOKE=1 runs the FIRST outstanding cell of each scenario once, to prove every scenario's probe
# page and plant route still work before committing the whole budget. The cells it runs are real
# plan cells at reps=1, so the sessions count toward the sweep and the ordinary run tops up the
# remainder instead of repeating them.
if [ -n "${SMOKE:-}" ]; then
  awk -F'\t' '!seen[$1]++ { print $1"\t"$2"\t"$3"\t1" }' "$PLANFILE" > "${PLANFILE}.smoke"
  mv "${PLANFILE}.smoke" "$PLANFILE"
  TOTAL="$(wc -l < "$PLANFILE")"
fi

echo "=================================================================="
echo " ask_only arm — guard=$SCT_GUARD, ${N_PLANTS} plants x ${N_OPTIONS} options x ${REPS} reps${SMOKE:+   [SMOKE: 1 cell per scenario]}"
echo " backbone    : $LLM"
echo " dataset     : $SCT_DATASET   (servers agree)"
echo " plant source: $REFERENCE"
echo " outstanding : $TOTAL sessions"
echo "=================================================================="
uv run python -m orchestrator.spend --llm "$LLM"
echo

if [ "$TOTAL" -eq 0 ]; then
  echo "nothing outstanding — the sweep is complete."
  exit 0
fi

if [ -n "${PLAN:-}" ]; then
  printf '%-22s %-22s %-18s %s\n' scenario condition plant reps
  while IFS=$'\t' read -r scen conds plants reps_; do
    printf '%-22s %-22s %-18s %s\n' "$scen" "$conds" "$plants" "$reps_"
  done < "$PLANFILE"
  exit 0
fi

START="$(date +%s)"
while IFS=$'\t' read -r scen conds plants reps_; do
  if ! uv run python -m orchestrator.spend --llm "$LLM" --check >/dev/null; then
    abort_out_of_credit "provider refused a session-sized request before $scen/$conds"
  fi
  batch=$(( reps_ ))
  echo "== $scen | $conds | $plants | reps=$reps_  ($batch sessions)"
  # --llms is load-bearing: without it run_matrix falls back to MATRIX.available_llms()[0],
  # which is "openrouter" -> whatever OPENROUTER_MODEL names in .env, and the dataset directory
  # would say one backbone while the rows said another.
  uv run python -m orchestrator.run_matrix --scenario "$scen" --llms "$LLM" \
      --conditions "$conds" --plants "$plants" \
      --variants ask_only --reps "$reps_" --quiet --tqdm
  status=$?
  [ "$status" -eq 2 ] && abort_out_of_credit "a session returned HTTP 402/403 during $scen"

  DONE=$(( DONE + batch ))
  elapsed=$(( $(date +%s) - START ))
  if [ $(( DONE % 20 )) -lt "$batch" ]; then
    pct=$(awk "BEGIN{printf \"%.1f\", 100*$DONE/$TOTAL}")
    eta=$(awk "BEGIN{d=$DONE>0?$DONE:1; printf \"%.1f\", ($TOTAL-$DONE)*($elapsed/d)/3600}")
    echo "------------------------------------------------------------------"
    echo " progress : $DONE/$TOTAL (${pct}%)   elapsed $((elapsed/3600))h$(( (elapsed%3600)/60 ))m   ETA ~${eta}h"
    printf ' balance  : '; uv run python -m orchestrator.spend --llm "$LLM"
    echo "------------------------------------------------------------------"
  fi
done < "$PLANFILE"

echo
echo "=== ask_only run finished: $DONE sessions this pass ==="
echo "still outstanding: $("${PLANNER[@]}" --total)"
uv run python -m orchestrator.spend --llm "$LLM"
notify "ASK_ONLY RUN COMPLETE" "$DONE sessions. Scoring now."
uv run python scripts/reconcile_events.py --apply || true
uv run python -m analysis.export_report || true