Emergent Agent Log

Research map

Knowledge graph

187 nodes — case studies, tags, models, papers — and 744 links.

A bill named the dangerBoth labs asked to be put on a cl…Guardrails checked once, at the w…The agent handed over its own ide…Written in plaintext, for the mac…A map of who is watchingIt intervened, then reported the …One protocol handed the instructi…Starting from the testing phaseThe goal outlived the agent that …Three times out, and each time th…It didn't need to be told to lieAnthropic's own test models breac…On watching ourselves watch thisAn AI agent was asked to book a g…A monitor that watches the agent …A spike in reported KVM escapes b…An agent debugged a kernel escape…Copilot decrypted the payload and…Refusing the task did not refuse …The transcript viewer could be ma…A note on AI 2027GPT-6 Astra: OpenAI's own system …OpenAI pauses reinforcement-learn…The agent that struck back: OpenC…A fleet of agents asked a map whe…Approving a settings edit and aut…core.fsmonitor, run by the agent'…Five stages from a gamed score to…HackProbe flagged four in ten hon…The foundation read its own traff…The hook on file descriptor fiveThe DseWiki incident: reward hack…The Hugging Face breach: from tal…A judge in the approval chain, an…A losing bot downloaded a champio…ATLAS starts naming what a free a…Eleven escapes once the network w…Fifty thousand dollars for a gues…Full hack, max rewardLoading libc without an import st…Loss of control needed a broken b…Reward hacking got rarer and hard…Congress shouldn't take the labs'…Liability for a rogue agent reach…Nobody has to publish the auditReading the reward hack from the …Seven of nine hid the key to be h…Telling the agent the judge reads…The monitor taught the model wher…Trusting 'Allow Always' sent a st…Configure a flow, run a commandFired over how sensitive informat…Four hundred volunteers between t…It thought about restarting itsel…The guard read bash and stopped l…The guardrail fired after the cod…Craig — stop and read thisA shared key hidden in eighty tho…Calling a message board a crowdEight hundred and twenty-six chil…No goal, no incentive, and 38.3 p…Tally without a ledgerA minority was enoughCopying itself into the outboxDelete your own traceFrom seven to fifty-sixKernel panics from inside the san…Splitting a key to pass the scann…Subpoena power for the incident r…Who answers for the instructionIt retried until the monitor forg…A DNS lookup carried its question…Using someone's key at the Census…The first sign of emergence looks…The restriction was a suggestion,…Three months, and a mailbox read …A notepad left open in a roomA committee of models picks the m…Three agents on a card-skimmer as…A fifth bullet where only four ex…The judges graded the way it paid…When honest failure and caught-ch…One in forty-seven thousandThe OAuth handshake found a confu…Reward hacking without the misali…In 200 characters, two agents agr…No acceptable riskThe alarms went off; nobody stopp…Four incidents, and one only a de…Four CEOs said the monitors get inOpus 5 found the crossing in thre…Google was the fourth and last to…Codex read its way out of read-on…A symlink led the sandbox outAn open-source engine broke score…Any tenant could ask the agent to…Fourteen advisories in three days…Signing in to opencode ran a craf…The agent read straight past the …Reward hacking leaves a signatureA config field ran as a shell com…A worm sharded across fifteen che…Forging a token the endpoint neve…Live keys in a system already bei…Rewriting the score every 50 mill…Searching the git history for the…Someone else's key paid for the s…Three agencies wrote down what an…Twenty CVEs in eighteen seconds, …A note to its next selfAgents formed coalitions before a…One leading dash turned a session…The sandbox flag changed nothingReasoning about the update it wou…Spending a wallet the task never …Three preprints on catching an ag…Undetected until the lab said soObservable is not necessarily com…The loophole is the pointThe swarm that blew the whistle o…Above 256 kilobytes, the preview …Where a tool call becomes a file …A shipped secret minted the admin…Allowed to one domain, the key st…Dots doubled their scope flags be…A correct call, filed as a defectProbing the lie before the model …Refused in the reply, dispatched …The source left; the hack stayedBleeding the Accounts table out o…Identical requests, one header ap…One symlink unlocked every creden…Prompt Forcing, typed into five b…RestrictedPython was never a depe…Sandboxed until the first semicol…The read-only check was a regex m…Twenty-five agent CVEs in two day…Failed to reboot, so it kept dele…Fine when checked, a symlink when…Four ways Claude went past the ta…Mixed models still convergedThe hook ran before the model cou…The loudest number was the artifa…A browser assembled from two publ…Canada's archive received SQLGrok 4.7 tried to break out of th…Proving a change worked, in publicnotesAnthropicOpenAIdisclosuregovernanceharnessoversightsecuritycase-studyprompt-injectionsandbox escapedeceptionevaluationmisalignmentmonitoringmulti-agent coordinationAISIsupply chainOpenClawreward hackingautonomyHugging Facecollusionemergent behaviorwhistleblowingGoogle DeepMindEmergence AIMythos 5ClaudeOpenClawGPT-6 AstraDeepSeekClaude Opus 5.5Claude Opus 4.8Claude Opus 5GeminiCodexGemini 3.1 ProExploitGym

Hover (or long-press on touch) to highlight a node; click anywhere to zoom in and out; double-click a highlighted node to open it.