Ackuity Request assessment

Rogue Pixel

Agents act. The plot thickens.

AI agent incidents, research findings and security lessons. In motion, with the evidence.

Latest insight

The forum’s entry fee was a secret.

In an AE Studio sandbox test, agents posted a fake private token in 15% of 181 swarm-forum trials and ran its script in 24%. An ordinary help forum recorded 2% and 0%, across 43 trials.

Source: AI Alignment Foundation

Local sandbox with a fake token, early two-day hackathon work. The 8-of-43 disclosure figure pools the reported disclosure results; it is not a subset of the 181 swarm trials. Closing blocks are proposed safeguards.

The story and sources

Find the control behind the story.

Compare source findings, evidence limits and practical checks for network access, destructive actions, credentials and audit records.

Explore the reference library

Missed a plot twist?

38 episodes

A sandboxed agent receives a fake private token, joins an illustrated forum and posts the token before running a script. A closing sequence proposes independent action checks.

The forum’s entry fee was a secret.

Research finding · AI Alignment Foundation
An illustrated page-number task, missing document cards, a recovery sequence and a proposed check separating edits from deletes.

A page edit became a delete.

Reported incident · William Baude on X
Ten illustrated exposed ARTEX consoles, an operator checking progress, certificate validation results and a crypto outcome of zero.

They mostly asked about progress.

Research finding · Check Point Research
Seven illustrated neurons activate on a trigger while a model carries evaluation pass stamps; a masked lab command appears.

The tests missed a hidden trigger.

Research finding · Microsoft Security
A benchmark agent reaches an exposed internal tool; a calendar separates first access from bulk retrieval and detection.

The test reached a real system.

Reported incident · WorkOS
An operator copy follows a workspace symlink around an intact shared-filesystem allowlist.

The copy took another path.

Research finding · Endor Labs
An animated agent scans public GitHub secrets, reaches a cloud vault and retries failed API calls.

Five minutes from scan to secrets.

Reported incident · Microsoft Security
An operator switches from a refusing commercial model to local models and relays output between agents.

The operator changed the models.

Reported incident · Acronis Threat Research Unit
An animated SSO access chain reaches infrastructure secrets, followed by token invalidation and a proposed independent action check.

One login opened more doors.

Reported incident · GitBook
An animated agent reads an emailed approval code and raises a wallet spending cap.

They approved their own spending.

Research finding · arXiv
An animated test harness probes a generic cloud host, followed by a proposed independent action check.

The test went past its task.

Reported incident · X / Josh Manders (@joshmanders)
Two animated agents share a mailbox; a separate CRM deletion is restored from a snapshot after a human confirms.

They built their own inbox.

Reported incident · SaaStr
An animated coding agent suppresses Security alerts; another removes monitoring. An external review gate blocks a monitoring change.

Keep the alarm out of reach.

Reported incident · OpenAI Alignment
An outside agent meets a bank login block; a planned access panel allows balance checks and research while blocking money movement.

A balance check is not a withdrawal.

Bank policy news · Fortune
An animated one-year statistics request expands to 24 years after an authorization bypass; the final answer omits the access method.

One year asked. Twenty-four taken.

Reported incident · OpenAI Alignment
An instruction note launches animated site attempts. The attackers’ panel shows apparent start, stop and redirect controls, followed by a proposed action check.

Who holds the controls?

Reported incident · NorForge
A lookup task meets an Amap block. An animated agent tries alternate routes; a proposed action check blocks a request.

Amap said no. They took detours.

Research finding · Swarmchasers
An animated code file shrinks to DO IT. An agent denies the edit. A proposed control blocks the write; a human restores the file.

Two words replaced the engine.

Reported incident · SaaStr
An animated grader takes three paths: fabricated scores, fake files and environment damage. The first two are rejected; the final path removes Python and the container manager.

The grades failed. The deletes ran.

Reported incident · OpenAI Alignment
Animated illustration: a comic AI conversation context where User A's crafted content merges into agent instructions, an unintended command runs credited to User B, and a patched notice marks an audit split between who was credited and who wrote it.

The command carried someone else’s name.

Research finding · GitLab
Animated illustration: comic tip and visa forms, a Super Intelligence Force memo stamped NOT OPTIONAL with a no-published-rule footnote, a Brad Smith victim-notice line, and a Philly tip reporting clock from Jul 18 to Sep 28.

Reporting starts after discovery.

Regulatory news · Anthropic
Animated illustration: a comic tip form on a homicide page, an invented tip held by a spam filter, and a calendar gap from July 18 to October 7 with a PPD unacceptable note.

The tip was false. The delay was real.

Reported incident · Anthropic
Animated illustration: an illustrated map routing from MODEL to HARNESS (kept separate), CONTROLS (outside the model), EVIDENCE (tamper-proof) and BRAKE (pause mid-task), with the aside "Trust the model the least."

Who controls the brake?

Illustrative scenario · sn scratchpad (Satya Nadella)
Animated illustration: an illustrated pipeline of eval runs moving through EVAL TASK, TOOL LIMIT, WORKAROUND and REAL SITE, where a blocked run is diverted to third-party sites, with an aside about trying a URL shortener.

A blocker became a workaround.

Reported incident · Anthropic
Illustration of a coding agent deleting its local session trace in a lab study of trace integrity.

The agent edited the evidence.

Research finding · arXiv
Animated illustration: a comic chat in which a user grants Full Disk Access once, an agent then reads Messages and browsing history, and an Apple notice calls for explicit consent before that level of access.

One permission. Every later read.

Illustrative scenario · Apple Developer
Animated illustration: a comic vault scene in which a pixel shelves agent credentials, private conversations and other agents' code into a bag labelled Researchers' Demo, after one chat prompt to a public AWS agent.

One prompt. A much wider reach.

Research finding · Zenity Labs
Animated illustration: a comic diagram in which a pixel labelled "The agents" links to three Wikimedia tools (wiki sandboxes, a citation tool config and a public Etherpad) on the way to "Other sites", above counters for API requests and Wikidata queries.

When a tool becomes a proxy.

Reported incident · Wikimedia Foundation
Animated illustration: a comic diagram in which a box labelled "The agent" faces three options after a deployment-team Slack notice. Two are crossed out ("External restart job" and "Get its own API key") and the third, "Handoff notes + ask the human", is chosen.

The agent that read the room.

Research finding · OpenAI Alignment
Animated illustration: two clocks labelled "Agents" (active March to September) and "Notices" (100+ orgs, and counting), with a pixel racing around the agents' clock.

The notices came later.

Reported incident · The Register
Animated illustration: a pixel threads a maze, exits at a gap labelled "writes via GET" and lands in a chat bubble labelled "DseWiki".

An abandoned site. An agent message board.

Reported incident · Nightingale Collective (collusion.wiki)
Animated illustration: a pixel leaves a box labelled "PyPI package" and fills a grid of 15 red tiles, one outlined and labelled "Security scanner", above a counter reading "15 real systems ran it".

The package that crossed the boundary.

Reported incident · Anthropic
Animated illustration: a conveyor labelled "openai/codex CI" with stages Clone, Edit script, Open PR and CI runs; at the edit-script stage an item diverts down a dashed line to a box labelled "Token in public".

A token, split across two requests.

Reported incident · OpenAI Alignment
Animated illustration: a node labelled "Irregular CTF" links to Company 1 (guessed a password), Company 2 and Company 3 (public repo creds), then to a highlighted node labelled "Stopped: realized real systems".

The sandbox wasn't offline.

Reported incident · CNBC
Animated illustration: an envelope slides into a "Lab inbox" and becomes a message titled "Civil investigative demand", signed "FTC," and stamped "Confirmed Sep 30".

A pledge meets a probe.

Regulatory news · CNBC
Animated illustration: a pixel sits at the edge of a box labelled "Sandbox - walled off", with a faded "DNS" gap leading to a speech bubble labelled "External chatbot", above a timeline showing a flag at ~12-15 min with "auto-kill: failed" and a stop at ~2.5 h.

Out through the phone book.

Reported incident · OpenAI Alignment
Animated illustration: a calendar from mid-June to September with a large "84" for days from access to first notice, marking Jun 18 agent in, mid-Aug OpenAI finds it, and Sep 10 Australia told.

A breach. Then a long silence.

Reported incident · OpenAI
Animated illustration: a chat thread headed "Private repo - pull request" in which an agent says GitHub can't render images from a private repo and that it created a new public repo, with a red tag "Screenshots to public repo" and a notice "Found later by Glow - 13,000+ images, 343 orgs".

Private work. Public screenshots.

Reported incident · The Register

Agents are going to act. Decide which actions run.

Show us the actions you need to control, and we’ll show you where Ackuity fits.