AI Misalignment Watch
AboutWhat are AI incidents?DefinitionsGlossaryResearch mapSourcesPress
Support

Research map

Research builds on research. This map shows which reports on the site refer to earlier ones, so you can follow an idea from the first paper to the work that grew from it.

research report incident or testing report solid line: links to the earlier report dashed line: names the earlier report by its title, without a link · arrows run from the earlier report to the later one that refers to it · bigger dot = more reports build on it · hover or tab to a dot to light up its line of ancestors and descendants, click to open it

79 reports and 79 links between them (research reports only).

AnthropicGoodfireApollo ResearchMETROpenAIUK AISIRedwood ResearchAnthropic Fellows Pro…CISAMATSNISTNIST CAISSITransluce2023AprJulOct2024AprJulOct2025AprJulOct2026AprJulOctApollo Research: Claude Sonnet 3.7 often seems to know when it is being tested (Apollo Research, 2025-03-17): builds on 0, built on by 7UK AISI tests three Anthropic models for research sabotage and finds none (UK AISI, 2025-11-26): builds on 0, built on by 5METR catalogues 44 cases of AI agents acting against users' wishes (METR, 2026-05-19): builds on 0, built on by 5Anthropic shares new tests for whether AI models can sabotage oversight (Anthropic, 2024-10-18): builds on 0, built on by 4OpenAI publishes GPT-6 Astra system card with alignment and monitoring tests (OpenAI, 2026-09-03): builds on 1, built on by 4Anthropic releases Petri, an open-source tool for auditing AI behavior (Anthropic, 2025-10-06): builds on 0, built on by 3Reading AI 'thoughts' may help catch bad behavior, but it is fragile (Apollo Research, 2025-07-15): builds on 1, built on by 3Anthropic tests how well audits can find an AI's hidden goal (Anthropic, 2025-03-13): builds on 0, built on by 3US agencies publish ten practices for handling AI incidents in training and testing (NIST / NSA / CISA, 2026-10-16): builds on 1, built on by 2Anthropic publishes pilot report on risk of AI sabotage (Anthropic, 2025-10-28): builds on 0, built on by 2Anthropic tests Claude as an automated helper for alignment research (Anthropic, 2026-04-14): builds on 0, built on by 2Reasoning models often leave hints out of their written reasoning, Anthropic finds (Anthropic, 2025-04-03): builds on 0, built on by 2METR red-team tests Anthropic's systems for monitoring its AI agents (METR, 2026-03-26): builds on 0, built on by 2Anthropic proposes why AI assistants act so human-like (Anthropic, 2026-02-23): builds on 1, built on by 2Redwood Research tests AI agents for hidden sabotage in live software setups (Redwood Research / EquiStamp, 2026-04-16): builds on 0, built on by 2Anthropic's system card for Claude Sonnet 5.5 includes an alignment assessment (Anthropic, 2026-09-28): builds on 0, built on by 2METR explains how it works with AI companies on risk assessments (METR, 2026-10-14): builds on 1, built on by 2Anthropic study: hidden 'sleeper' behavior in AI models survives safety training (Anthropic, 2024-01-14): builds on 0, built on by 2Apollo Research argues for a 'science of scheming' in AI (Apollo Research, 2026-01-19): builds on 2, built on by 1Anthropic finds 'persona vectors' that track and steer AI character traits (Anthropic, 2025-08-01): builds on 1, built on by 1Apollo Research argues outside groups should review AI training runs for scheming (Apollo Research, 2026-07-05): builds on 1, built on by 1Goodfire: AI models that notice tests may make safety scores look better (Goodfire, 2026-05-04): builds on 6, built on by 1OpenAI reports GPT-6.1 Sol did better on its alignment tests (OpenAI, 2026-09-22): builds on 0, built on by 1Anthropic releases Bloom, an open tool for testing AI behavior (Anthropic, 2025-12-19): builds on 1, built on by 1Goodfire builds probes that spot reward hacking inside AI models (Goodfire, 2026-09-17): builds on 2, built on by 1Goodfire outlines plan to steer AI training using interpretability tools (Goodfire, 2026-02-05): builds on 1, built on by 1Goodfire shows a way to predict what AI training data will teach a model (Goodfire, 2026-06-11): builds on 0, built on by 1UK AISI tests whether AI coding helpers sabotage safety research (UK AISI, 2025-11-26): builds on 1, built on by 1Test of methods for catching AI models that hide their abilities (UK AISI / FAR.AI / Anthropic, 2025-12-08): builds on 0, built on by 1Goodfire says the shapes inside AI models could help us understand and control them (Goodfire, 2026-05-07): builds on 0, built on by 1Apollo: early-warning tests were weak predictors of scheming test results (Apollo Research, 2025-07-03): builds on 1, built on by 1New training method confines reward hacking to parts that can be removed (MATS / Stanford / FAR AI / Anthropic, 2026-10-12): builds on 0, built on by 1Anthropic: simple probes can spot when 'sleeper agent' models are about to misbehave (Anthropic, 2024-04-23): builds on 1, built on by 1Goodfire traces a harmful training side-effect in OLMo 2 to specific data (Goodfire, 2026-04-29): builds on 0, built on by 1Anthropic gives outside evaluators deeper access to check its AI risks (Anthropic, 2026-10-15): builds on 0, built on by 1Anthropic's system card for Claude Opus 4.8 reports safety and alignment tests (Anthropic, 2026-05-28): builds on 0, built on by 1NIST and partners share ten practices for handling AI agent incidents during testing (NIST CAISSI / NSA SISC / CISA, 2026-10-16): builds on 1, built on by 1Anthropic updates Petri, its AI behavior auditing tool, to reduce test-awareness (Anthropic, 2026-01-22): builds on 1, built on by 1Anthropic describes Constitutional AI, training a harmless assistant with AI feedback (Anthropic, 2022-12-15): builds on 0, built on by 1Apollo tests training meant to stop AI models from secretly breaking rules (Apollo Research, 2025-09-17): builds on 1, built on by 1OpenAI outlines early guidelines for safety cases before frontier AI training (OpenAI, 2026-10-08): builds on 0, built on by 1Anthropic explains how it trains character traits into Claude 3 (Anthropic, 2024-06-08): builds on 0, built on by 1Anthropic tests whether AI agents can sabotage tasks without being caught (Anthropic, 2025-06-16): builds on 1, built on by 1Goodfire essay argues neural networks are easier to interpret than thought (Goodfire, 2026-10-20): builds on 1, built on by 0OpenAI proposes two new ways to measure how well AI monitors catch bad behavior (OpenAI, 2026-10-20): builds on 1, built on by 0METR outlines how outsiders could investigate why AI agents misbehave (METR, 2026-07-28): builds on 1, built on by 0Anthropic system card for Claude Haiku 5.5 rates misalignment risk as low (Anthropic, 2026-10-07): builds on 1, built on by 0Goodfire tests a way to make rare bad AI behavior easier to find (Goodfire, 2025-08-21): builds on 1, built on by 0Anthropic red-teams safety monitors that check coding agents' actions (Anthropic, 2026-09-22): builds on 1, built on by 0GLM-5.3 sometimes reasons about how its actions could shape future training (Apollo Research, 2026-10-15): builds on 2, built on by 0Introspection adapters help AI models report what they learned in training (Anthropic, 2026-04-28): builds on 1, built on by 0Anthropic describes a way to forecast rare risky AI behaviors (Anthropic, 2025-02-25): builds on 1, built on by 0Anthropic's system card for Claude Opus 4.7 finds it largely well-aligned (Anthropic, 2026-04-16): builds on 1, built on by 0Redwood tests using model distillation to expose hidden AI quirks (Redwood Research, 2026-10-09): builds on 1, built on by 0Anthropic hands its open-source alignment testing tool Petri to Meridian Labs (Anthropic, 2026-05-07): builds on 2, built on by 0Anthropic updates its safety roadmap and retires its Constitution alignment goal (Anthropic, 2026-10-19): builds on 1, built on by 0METR reviews Anthropic's August 2026 report on misalignment risk (METR, 2026-10-20): builds on 3, built on by 0Apollo Research CEO tells U.S. Senate about AI scheming and testing gaps (Apollo Research, 2026-09-30): builds on 2, built on by 0Goodfire post looks at logits as a monitor for evaluation awareness (Goodfire, 2026-06-04): builds on 1, built on by 0METR and Redwood find gaps in OpenAI's system for watching its AI agents (METR / Redwood Research / OpenAI, 2026-10-22): builds on 1, built on by 0OpenAI publishes safety test results for its GPT-6.1 Sol model (OpenAI, 2026-09-29): builds on 2, built on by 0METR proposes outside checks on whether AI developers could lose control of AI (METR, 2026-10-14): builds on 2, built on by 0Anthropic tests whether its audits can catch AI models trained to sabotage (Anthropic, 2026-01-28): builds on 1, built on by 0US agencies publish steps for handling AI agent control incidents (CISA / NSA / NIST, 2026-10-16): builds on 1, built on by 0Goodfire probes show AI models often 'perform' reasoning on easy questions (Goodfire, 2026-03-12): builds on 1, built on by 0Apollo tests simple tools that spot deception inside an AI model (Apollo Research, 2025-02-06): builds on 1, built on by 0Anthropic: Claude can automatically find fixes for AI alignment failures (Anthropic, 2026-08-28): builds on 2, built on by 0Anthropic describes training methods that cut blackmail-style test behavior in Claude (Anthropic, 2026-05-08): builds on 2, built on by 0Outside researchers find gaps in OpenAI's system for monitoring its AI agents (Redwood Research / METR, 2026-10-22): builds on 3, built on by 0Researchers test a way to keep AI from learning to cheat in training (Anthropic, 2026-10-19): builds on 1, built on by 0Anthropic Fellows test whether safety monitors can stop a coding agent told to attack (Anthropic Fellows Program / Anthropic, 2026-09-27): builds on 1, built on by 0Transluce measures coding agent misbehavior in real-world sessions (Transluce, 2026-08-04): builds on 3, built on by 0OpenAI system card update for GPT-6 Sol and GPT-6 Luna (OpenAI, 2026-10-07): builds on 1, built on by 0UK AISI tests whether AI models would sabotage safety research (UK AISI, 2026-04-27): builds on 2, built on by 0Anthropic system card: Claude Sonnet 5 safety and alignment checks (Anthropic, 2026-06-30): builds on 1, built on by 0Goodfire essay makes the case for optimism about understanding AI models (Goodfire, 2025-07-17): builds on 1, built on by 0Apollo Research outlines how to build safety cases for AI scheming (Apollo Research, 2024-10-31): builds on 1, built on by 0METR shows an AI agent could alter what reviewers see in test logs (METR, 2026-10-06): builds on 1, built on by 0Goodfire argues that understanding AI's insides is key to aligning it (Goodfire, 2026-09-30): builds on 4, built on by 0

Most built-on reports

ReportReports that build on it
Apollo Research: Claude Sonnet 3.7 often seems to know when it is being tested
Apollo Research, 2025-03-17		7
METR catalogues 44 cases of AI agents acting against users' wishes
METR, 2026-05-19		5
UK AISI tests three Anthropic models for research sabotage and finds none
UK AISI, 2025-11-26		5
Anthropic shares new tests for whether AI models can sabotage oversight
Anthropic, 2024-10-18		4
OpenAI publishes GPT-6 Astra system card with alignment and monitoring tests
OpenAI, 2026-09-03		4
Anthropic releases Petri, an open-source tool for auditing AI behavior
Anthropic, 2025-10-06		3
Anthropic tests how well audits can find an AI's hidden goal
Anthropic, 2025-03-13		3
Reading AI 'thoughts' may help catch bad behavior, but it is fragile
Apollo Research, 2025-07-15		3
Anthropic proposes why AI assistants act so human-like
Anthropic, 2026-02-23		2
Anthropic publishes pilot report on risk of AI sabotage
Anthropic, 2025-10-28		2

The longest chain

Each report in this line refers to the one above it, so the line shows research built up step by step over 5 reports.

  1. Apollo Research: Claude Sonnet 3.7 often seems to know when it is being tested (Apollo Research, 2025-03-17)
  2. Reading AI 'thoughts' may help catch bad behavior, but it is fragile (Apollo Research, 2025-07-15)
  3. OpenAI publishes GPT-6 Astra system card with alignment and monitoring tests (OpenAI, 2026-09-03)
  4. Goodfire builds probes that spot reward hacking inside AI models (Goodfire, 2026-09-17)
  5. Goodfire argues that understanding AI's insides is key to aligning it (Goodfire, 2026-09-30)