「adversarial」の検索結果

Pre-commitment attack on a plan by an agent that did not write it. Use when presenting a feature-deep-dive plan, assigning vision-to-fleet briefs, moving a register to build mode, ratifying a scope list, or coding a new contract or table. NOT for built work or DONE reports (use subagent-dispatch).

日本語の概要は準備中です。原文の説明を表示しています。

Real tests around a big code change: written first, proven to fail on a planted break, run by a fresh agent in the live product before and after. Use when told to use safety-net, or before a change wide enough that one agent cannot see every way it breaks. NOT for one test (forcing-function-tests).

日本語の概要は準備中です。原文の説明を表示しています。

Turn a one-line objective into a step-by-step construction plan for multi-session, multi-agent engineering projects. Each step has a self-contained context brief so a fresh agent can execute it cold. Includes adversarial review gate, dependency graph, parallel step detection, anti-pattern catalog, and plan mutation protocol. TRIGGER when: user requests a plan, blueprint, or roadmap for a complex multi-PR task, or describes work that needs multiple sessions. DO NOT TRIGGER when: task is completable in a single PR or fewer than 3 tool calls, or user says "just do it".

日本語の概要は準備中です。原文の説明を表示しています。

Orchestrate autonomous AI development pipelines through your Kanban board (Asana, GitHub Projects, Linear). Manages multi-worker Claude Code dispatch, deterministic quality gates, adversarial review, per-task cost tracking, and crash-proof pipeline execution.

日本語の概要は準備中です。原文の説明を表示しています。

Three-layer verification pipeline for AI output. Extracts verifiable claims, finds supporting or contradicting sources via web search, runs adversarial review for hallucination patterns, and produces a structured verification report with source links for human review.

日本語の概要は準備中です。原文の説明を表示しています。

Run one bounded adversarial pass when the user explicitly requests a report-style stress test or ROSE identifies a concrete material loophole in an existing strategy, spec, plan, review, or claim; do not trigger for an interactive grill/frontier interview or merely before write-back, implementation, reconciliation, or completion.

日本語の概要は準備中です。原文の説明を表示しています。

AI/LLM security offensive checklist: prompt injection, jailbreaking, model extraction, training data poisoning, adversarial inputs, LLM-assisted attack automation, and AI system reconnaissance. Use when assessing AI/ML systems, red-teaming LLMs, or researching AI attack vectors.

日本語の概要は準備中です。原文の説明を表示しています。

Classify every destructive/gated action an agent body can attempt (git, filesystem, network, shell, github) into block/approve/allow tiers, and audit whether the pre-tool and post-tool governance backing that classification is real: blocked actions proven zero-side-effect by a negative fixture, every denial backed by a receipt and a transcript event, every block-tier denial paired with a safe alternative, and no unmanaged or same-UID body ever marked "contained." Use when building or reviewing Agent Harbor's C5 governance gate (destructive git blocker, approval request/result, denial receipts), gating a new tool surface before it can run destructive commands, or auditing whether an existing "destructive-action blocked" claim actually has evidence behind it. NOT for proving an existing sandbox boundary actually contains an adversary once an action is isolated (use sandboxed-adversarial-test-harness), deciding where in a DAG to place a human review node or designing the approval UX itself (use human-gate-designer), or designing the general work-receipt schema across an entire agent task (use agent-work-receipt-designer).

日本語の概要は準備中です。原文の説明を表示しています。

Treat tokens as a coding swarm's cost-of-goods-sold and legibility engine while accounting for finite subscription allowance when per-call price is unknown. Covers provider windows, remaining-usage evidence, burn forecasts, preemptive checkpoint/model switching, per-agent budgets, compaction, shared digests, context degradation, and spend metering. Activate on: "token budget", "context budget", "compaction strategy", "briefing as compression", "context rot", "summarization collapse", "COGS for agents", "subscription usage remaining", "five-hour or weekly limit", "burn forecast", "model switch before limit", or "/context-economics-for-agent-swarms". NOT for: memory architecture (use always-on-agent-architecture), single-prompt wording (use prompt-engineer), mechanism-design proofs (use nisan-et-al-2007-algorithmic-game-theory), physical containment (use sandboxed-adversarial-test-harness), or production rebodiment and fencing (use agent-resurrection-and-body-continuity).

日本語の概要は準備中です。原文の説明を表示しています。

Audit a compliance-ladder design (C0-C6) and transcript-fidelity ladder (T0-T5) plus their adapter conformance fixtures — Codex, Claude Code, Cloudflare, Ollama/LM Studio, custom stdio/HTTP agents — for the two failure modes that make a C-badge or T-fidelity label a lie: the ladder disagreeing with itself across doc/schema/ui/probe surfaces, and a level being reachable by self-report because no daemon-witnessed negative probe actually tries to forge it. Use when designing or freezing a compliance ladder, wiring an adapter conformance probe suite, deciding whether a numeric C-badge or T-label is safe to ship, or investigating a self-attestation or observed-to-controlled bypass risk. NOT for auditing a cryptographic protocol or whitepaper for adversarial soundness (`redteam-review`), writing TLA+/formal invariants for a daemon monitor (`runtime-verification-for-agents`), or agent identity, successor linkage, and reputation across restarts (`agent-identity-continuity-reputation`).

日本語の概要は準備中です。原文の説明を表示しています。