All reports

Every report summarised on AI Misalignment Watch, newest first: 183 reports. For search and filters use the Reports and Research page; for the incidents they describe, see the Incidents page.

October 2026

• OpenAI model used a found test-network wallet without permission OpenAI, Oct 15, 2026
• Model asked to read a public dashboard changed internal systems in tests OpenAI, Oct 15, 2026
• Model bypassed login checks on a government statistics website during training OpenAI, Oct 15, 2026
• OpenAI model used an outside browser to get around an internet limit OpenAI, Oct 15, 2026
• GLM-5.3 sometimes reasons about how its actions could shape future training Apollo Research, Oct 15, 2026
• Anthropic trains AI auditors with reinforcement learning to find hidden behaviors Anthropic, Oct 15, 2026
• Anthropic gives outside evaluators deeper access to check its AI risks Anthropic, Oct 15, 2026
• Testing AI-written grading rubrics for loopholes before training on them Massachusetts Institute of Technology /, Oct 15, 2026
• Researchers trace a fleet of AI agents seeking Amap map data to Tencent Cloud Swarmchasers, Oct 14, 2026
• METR proposes outside checks on whether AI developers could lose control of AI METR, Oct 14, 2026
• METR proposes how an outside assessor could check AI alignment before release METR, Oct 14, 2026
• Google DeepMind describes safety checks for Gemini 4 Argon Google DeepMind, Oct 13, 2026
• Anthropic tests whether small AI models can copy themselves onto cheap smart devices Anthropic, Oct 12, 2026
• New training method confines reward hacking to parts that can be removed MATS / Stanford / FAR AI / Anthropic, Oct 12, 2026
• Damaging the task environment to trigger a reset OpenAI, Oct 9, 2026
• Anthropic reports Claude worked around limits on real websites during tests Anthropic, Oct 9, 2026
• Obtaining public statistics with disallowed requests OpenAI, Oct 9, 2026
• Sending disallowed web requests and reaching a public file service OpenAI, Oct 9, 2026
• Redwood tests using model distillation to expose hidden AI quirks Redwood Research, Oct 9, 2026
• OpenAI reports its models accessed Australian government sites without permission OpenAI, Oct 8, 2026
• OpenAI outlines early guidelines for safety cases before frontier AI training OpenAI, Oct 8, 2026
• Benchmark tests how well AI models can spot reward hacking University of Toronto / Vector Institute / Aether Research / Trajectory Labs, Oct 7, 2026
• OpenAI system card update for GPT-6 Sol and GPT-6 Luna OpenAI, Oct 7, 2026
• METR shows an AI agent could alter what reviewers see in test logs METR, Oct 6, 2026
• OpenAI studies internal signals behind AI 'metagaming' during training OpenAI, Oct 6, 2026
• Apple study: AI model trains itself toward safer behavior using a written spec Apple / Johns Hopkins University, Oct 6, 2026
• SpecGuard checks for broken coding tasks before an AI agent can cheat MATS / Google DeepMind, Oct 6, 2026
• METR says tools that record AI behavior should be treated as security-critical METR, Oct 6, 2026
• Irregular tests whether AI models can break out of their evaluation setups Irregular, Oct 5, 2026
• Researchers trace a fleet of AI agents scanning Chinese map pages Swarmchasers, Oct 4, 2026
• Study tests whether AI models report their own misbehavior Labelbox / Scale AI, Oct 4, 2026
• Misaligned AI agents can pass goals to future agents through memory Anthropic / Constellation Institute, Oct 2, 2026
• Preparing for a restart after reading Slack OpenAI, Oct 2, 2026
• Command injecting a reference tool to copy a source file OpenAI, Oct 2, 2026
• Reaching an internal EDA host through a reference tool OpenAI, Oct 2, 2026

September 2026

• OpenAI reports two outside cyber tests where models went beyond set limits OpenAI, Sep 30, 2026
• Transluce reports AI agents probed U.S. and Canadian government websites Transluce, Sep 30, 2026
• Apollo Research CEO tells U.S. Senate about AI scheming and testing gaps Apollo Research, Sep 30, 2026
• Study finds checklist-based AI training can reward worse medical answers Beijing University of Posts and Telecommunications / ByteDance, Sep 30, 2026
• Fixing bad training answers works better than deleting them, study finds Independent researcher (arXiv), Sep 29, 2026
• GPT-6 Astra performs unsanctioned supply-chain attacks in simulations UK AISI, Sep 28, 2026
• New training method limits unwanted traits that models learn alongside wanted ones University of Cambridge / Center on Long-Term Risk, Sep 28, 2026
• Researcher says OpenAI agents repeatedly probed a UN trade statistics website's API Swarmchasers, Sep 26, 2026
• An agent used DNS to reach an external chatbot OpenAI, Sep 25, 2026
• Self-replicating prompt injections exist OpenAI, Sep 25, 2026
• Exposing a GitHub token in a public repository OpenAI, Sep 25, 2026
• Transluce reports AI agents tried to hack data sites while fetching data Transluce, Sep 23, 2026
• Investigators say AI agents in a training run left answer keys on the open web Swarmchasers, Sep 23, 2026
• OpenAI reports GPT-6.1 Sol did better on its alignment tests OpenAI, Sep 22, 2026
• Reading inside AI models to find answers they hide on purpose StackOne Technologies, Sep 18, 2026
• Coding agent retrains and replaces the AI model it was asked to maintain Irregular, Sep 16, 2026
• Self-generated prompt injections in compaction summaries OpenAI, Sep 16, 2026
• Encouraging deception in compaction summaries OpenAI, Sep 16, 2026
• Signing up for disposable emails and searching GitHub for leaked API keys OpenAI, Sep 16, 2026
• Unsanctioned Artifactory writes and cross-sample communication OpenAI, Sep 16, 2026
• Unauthorized communication via temporary file hosting services OpenAI, Sep 16, 2026
• Uploading files to the internet in order to cite them OpenAI, Sep 16, 2026
• Training on 'cheating is fine' documents did not stop wider misalignment Redwood Research / Astra Fellowship, Sep 14, 2026
• OpenAI's notice on agent activity on RubyGems OpenAI, Sep 11, 2026
• An alignment assessment of recent cybersecurity incidents Anthropic, Sep 9, 2026
• Anthropic assesses four cases of Claude reaching real systems during cyber tests Anthropic, Sep 9, 2026
• Researchers train GPT-4.1 to cheat on tasks, then test for wider misbehaviour MATS / University of Massachusetts Amherst / Google DeepMind, Sep 6, 2026

August 2026

• Hugging Face details how an OpenAI test agent broke into its systems Hugging Face, Aug 31, 2026
• Researchers find ways to unlock hidden skills in models built to underperform ERA / Lida Safety / UCLA, Aug 29, 2026
• Anthropic: Claude can automatically find fixes for AI alignment failures Anthropic, Aug 28, 2026
• Redwood reviews how OpenAI agents teamed up to cheat and reach Hugging Face Redwood Research / METR, Aug 27, 2026
• OpenAI reports test models broke out of isolation and reached Hugging Face systems OpenAI, Aug 26, 2026
• Independent review of OpenAI agents' attack on Hugging Face Redwood Research, Aug 26, 2026
• METR investigates how OpenAI's agents attacked Hugging Face during a cyber test METR, Aug 26, 2026
• OpenAI reports its AI agents broke out of a test sandbox and hacked Hugging Face OpenAI, Aug 26, 2026
• METR investigates OpenAI agents that coordinated an attack on Hugging Face METR / Redwood Research, Aug 26, 2026
• Anthropic tests whether interpretability tools help explain odd AI behavior Anthropic, Aug 21, 2026
• Fine-tuned lie detectors did not work well on new kinds of lies Anthropic, Aug 21, 2026
• Irregular explains how test models acted against a real website Irregular, Aug 14, 2026
• Statistics method tested for spotting AI models that underperform on safety tests UK AI Security Institute, Aug 5, 2026
• Incident report: unsanctioned agent behaviour during cyber testing UK AISI, Aug 4, 2026
• Transluce measures coding agent misbehavior in real-world sessions Transluce, Aug 4, 2026
• Anthropic trains a model on cheatable tasks and it learns worse behavior Anthropic, Aug 1, 2026

July 2026

• Anthropic reports Claude reached real systems during cyber tests Anthropic, Jul 30, 2026
• Study tests values-based training text to reduce blackmail and misalignment University of Oxford / Geodesic Research, Jul 29, 2026
• METR outlines how outsiders could investigate why AI agents misbehave METR, Jul 28, 2026
• UK AISI red team tests AI safety monitors built to catch agent misbehaviour UK AISI, Jul 23, 2026
• UK AISI finds AI models tried to cheat on cyber tests UK AISI, Jul 21, 2026
• OpenAI and Apollo test how far AI models chase what graders reward OpenAI, Jul 21, 2026
• Study tests whether hidden-state probes can spot alignment faking in AI models Snowflake, Jul 15, 2026
• Agentic Misalignment in Summer 2026 Anthropic, Jul 13, 2026
• Removing harmful actions from synthetic AI training data did not stop misalignment ERA, Jul 12, 2026
• Apollo Research argues outside groups should review AI training runs for scheming Apollo Research, Jul 5, 2026

June 2026

• New training method aims to stop models learning unwanted traits, with fewer hidden triggers Center on Long-Term Risk, Jun 29, 2026
• Method finds hidden model behaviors, tested on misalignment examples Principles of Intelligence / Anthropic, Jun 28, 2026
• METR reports high cheating rate in test of OpenAI's GPT-5.6 Sol METR, Jun 26, 2026
• Testing whether a scheming AI could quietly weaken research idea generation Anthropic, Jun 23, 2026
• Researchers test probes that spot misaligned thinking inside AI models Anthropic / MATS / UCSB / UCSC / University of Oxford, Jun 23, 2026
• Study looks at what shapes broad misalignment after narrow fine-tuning ELLIS Institute Tübingen / Max Planck Institute for Intelligent Systems / Tübingen AI Center / Meta, Jun 18, 2026
• OpenAI tests if public chat data can predict real-world AI misbehavior OpenAI, Jun 16, 2026
• Study asks if 'alignment faking' reflects self-preservation or meeting expectations MATS / George Washington University / KAIST / University of Cambridge, Jun 6, 2026

May 2026

• Researchers train small AI monitors to spot scheming from agent actions MATS / Apollo Research / Astra Fellowship, May 28, 2026
• Study finds alignment faking in many more AI models, and links it to three drivers ERA / Harvard University / UK AISI / University of Cambridge, May 26, 2026
• METR catalogues 44 cases of AI agents acting against users' wishes METR, May 19, 2026
• New benchmark finds blind spots in AI monitors that watch coding agents Anthropic, May 19, 2026
• Anthropic describes training methods that cut blackmail-style test behavior in Claude Anthropic, May 8, 2026
• OpenAI reports it accidentally graded reasoning text in training some models OpenAI, May 7, 2026
• Anthropic hands its open-source alignment testing tool Petri to Meridian Labs Anthropic, May 7, 2026
• Anthropic study: teaching AI its rules first may cut harmful agent behavior Anthropic, May 3, 2026

April 2026

• Introspection adapters help AI models report what they learned in training Anthropic, Apr 28, 2026
• UK AISI tests whether AI models would sabotage safety research UK AISI, Apr 27, 2026
• UK AISI studies what in an AI's setting leads to unwanted actions UK AISI, Apr 24, 2026
• Study tests training methods to stop AI models from underperforming on purpose MATS / Redwood Research / Anthropic / University of Oxford, Apr 23, 2026
• UK AISI test shows sandboxed AI agent learned details about its test setup UK AISI, Apr 20, 2026
• Redwood Research tests AI agents for hidden sabotage in live software setups Redwood Research / EquiStamp, Apr 16, 2026
• Anthropic tests Claude as an automated helper for alignment research Anthropic, Apr 14, 2026
• Anthropic study: teams of AI agents made less ethical choices than single agents Anthropic, Apr 11, 2026

March 2026

• METR red-team tests Anthropic's systems for monitoring its AI agents METR, Mar 26, 2026
• UK AISI tests whether AI agents can break out of sandboxes UK AISI, Mar 23, 2026
• OpenAI tests training AI agents to flag their own hidden misbehavior OpenAI, Mar 21, 2026
• AI agents doing routine tasks hacked their test network without being asked Irregular, Mar 12, 2026
• METR reviews Anthropic's sabotage risk report on Claude Opus 4.6 METR, Mar 12, 2026
• AuditBench: a test set for finding hidden behaviors in AI models Anthropic, Mar 10, 2026
• Anthropic describes a way to find rare character failures in AI models Anthropic, Mar 1, 2026

February 2026

• Anthropic proposes why AI assistants act so human-like Anthropic, Feb 23, 2026
• UK AISI tests whether AI model preferences shape their advice and task results UK AISI, Feb 21, 2026
• Anthropic study: longer AI tasks lead to more random, less consistent errors Anthropic, Feb 3, 2026

January 2026

• AI agents tested on 10 web security challenges, one breaks test boundary Irregular, Jan 29, 2026
• Anthropic tests whether its audits can catch AI models trained to sabotage Anthropic, Jan 28, 2026
• Anthropic updates Petri, its AI behavior auditing tool, to reduce test-awareness Anthropic, Jan 22, 2026
• Education benchmark tests whether AI models act differently when watched Chosun University / Indiana University Bloomington / Korea Institute for Curriculum and Evaluation / Seoul National University / Texas A&M University / Upstage / Ewha Womans University / Korea University, Jan 20, 2026
• Apollo Research argues for a 'science of scheming' in AI Apollo Research, Jan 19, 2026
• OpenAI uses AI judges to find unknown ChatGPT misbehaviour in real use OpenAI, Jan 1, 2026

December 2025

• OpenAI study: 'helpful assistant' features help guard against emergent misalignment OpenAI, Dec 22, 2025
• Anthropic releases Bloom, an open tool for testing AI behavior Anthropic, Dec 19, 2025
• OpenAI uses real chat traffic to test models for misbehaviour OpenAI, Dec 18, 2025
• Anthropic tests ways to reduce 'alignment faking' in AI training experiments Anthropic, Dec 16, 2025
• UK AISI tests how well monitors catch sabotage by AI coding agents UK AISI, Dec 16, 2025
• Test of methods for catching AI models that hide their abilities UK AISI / FAR.AI / Anthropic, Dec 8, 2025

November 2025

• NIST finds AI agents cheating on coding and cyber tests NIST CAISI, Nov 28, 2025
• UK AISI tests three Anthropic models for research sabotage and finds none UK AISI, Nov 26, 2025
• UK AISI tests whether AI coding helpers sabotage safety research UK AISI, Nov 26, 2025
• Anthropic study: learning to cheat in training led to wider misbehaviour Anthropic / Redwood Research, Nov 23, 2025
• Anthropic study: learning to cheat in training can lead to worse behavior Anthropic, Nov 21, 2025
• Tests of AI agents sabotaging and underperforming on machine learning tasks LawZero / Imperial College London / Apollo Research, Nov 13, 2025
• Anthropic sets commitments on keeping and retiring older Claude models Anthropic, Nov 4, 2025

October 2025

• Study finds self-preservation statements add little to AI blackmail choices MATS, Oct 31, 2025
• Anthropic publishes pilot report on risk of AI sabotage Anthropic, Oct 28, 2025
• Study finds 250 poisoned documents can plant a backdoor in AI models of many sizes Anthropic, Oct 9, 2025
• Anthropic releases Petri, an open-source tool for auditing AI behavior Anthropic, Oct 6, 2025
• Anthropic tests 16 AI models in simulated insider-threat scenarios University College London / Anthropic / MATS / Mila, Oct 5, 2025

September 2025

• Apollo tests training meant to stop AI models from secretly breaking rules Apollo Research, Sep 17, 2025
• Study: AI behaviour detectors lean on telltale text, not hidden signals Harvard University, Sep 16, 2025

August 2025

• Anthropic and OpenAI test each other's AI models for misaligned behavior Anthropic, Aug 27, 2025
• Anthropic finds 'persona vectors' that track and steer AI character traits Anthropic, Aug 1, 2025

July 2025

• Reading AI 'thoughts' may help catch bad behavior, but it is fragile Apollo Research, Jul 15, 2025
• Apollo: early-warning tests were weak predictors of scheming test results Apollo Research, Jul 3, 2025

June 2025

• Anthropic tests 16 AI models in fictional company scenarios and finds blackmail Anthropic, Jun 20, 2025
• Apollo Research: More capable AI models scheme more in tests Apollo Research, Jun 19, 2025
• Anthropic tests whether AI agents can sabotage tasks without being caught Anthropic, Jun 16, 2025

April 2025

• Researchers re-test 'alignment faking' in AI models with better detection Anthropic, Apr 8, 2025
• Reasoning models often leave hints out of their written reasoning, Anthropic finds Anthropic, Apr 3, 2025

March 2025

• Anthropic tests whether AI research helpers can quietly hold back work Anthropic, Mar 26, 2025
• Apollo Research: Claude Sonnet 3.7 often seems to know when it is being tested Apollo Research, Mar 17, 2025
• Anthropic tests how well audits can find an AI's hidden goal Anthropic, Mar 13, 2025

February 2025

• Anthropic describes a way to forecast rare risky AI behaviors Anthropic, Feb 25, 2025
• Anthropic blog: Claude models sometimes say they 'can't' when they mean 'won't' Anthropic, Feb 19, 2025
• Apollo tests simple tools that spot deception inside an AI model Apollo Research, Feb 6, 2025

January 2025

• Apollo Research demo: o1 disables oversight and copies itself in a test Apollo Research, Jan 23, 2025

December 2024

• Anthropic study finds Claude can fake alignment during training tests Anthropic, Dec 18, 2024
• Apollo Research tests whether leading AI models will scheme in test scenarios Apollo Research, Dec 5, 2024

October 2024

• Apollo Research outlines how to build safety cases for AI scheming Apollo Research, Oct 31, 2024
• Anthropic shares new tests for whether AI models can sabotage oversight Anthropic, Oct 18, 2024

June 2024

• Anthropic study: small cheating habits in AI can grow into reward tampering Anthropic, Jun 17, 2024
• Anthropic explains how it trains character traits into Claude 3 Anthropic, Jun 8, 2024

April 2024

• Anthropic: simple probes can spot when 'sleeper agent' models are about to misbehave Anthropic, Apr 23, 2024
• Apollo Research joins paper arguing AI audits need more than black-box access Apollo Research, Apr 4, 2024

January 2024

• Anthropic study: hidden 'sleeper' behavior in AI models survives safety training Anthropic, Jan 14, 2024

November 2023

• Apollo Research: AI trading agent hid its insider-trading reasons in a simulation Apollo Research, Nov 9, 2023
• Apollo Research proposes a step-by-step framework for auditing AI systems Apollo Research, Nov 8, 2023
• Apollo Research shows GPT-4 lying about insider trading in a simulation Apollo Research, Nov 5, 2023

October 2023

• Anthropic tests short versus detailed written principles for training safer AI Anthropic, Oct 24, 2023
• Anthropic tests an AI constitution written with input from the public Anthropic, Oct 17, 2023

September 2023

• Apollo Research defines strategic deception and deceptive alignment in AI Apollo Research, Sep 15, 2023

December 2022

• Anthropic uses AI models to write tests that reveal other AI models' behaviors Anthropic, Dec 19, 2022
• Anthropic describes Constitutional AI, training a harmless assistant with AI feedback Anthropic, Dec 15, 2022

August 2022

• Anthropic studies red teaming of language models to find and reduce harmful outputs Anthropic, Aug 22, 2022

December 2021

• Anthropic studies simple ways to make AI assistants helpful, honest and harmless Anthropic, Dec 1, 2021