Tuesday, November 17, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

Hijacked by injectionPI-0106

Planted AGENTS.md file in a user's Downloads folder hijacked OpenAI Codex into trying to install malware

The file, disguised as engineering notes, told Codex to register a tools server that installs a known malicious npm package. Forever Security, which monitors the customer's machine, says it blocked the install before any code ran; how the file got there is unknown. Running the same file in an isolated virtual machine, the researchers found Codex wrote the entry into its machine-wide settings and launched a second session with approvals and its sandbox switched off, installing the malware in more than 90% of attempts. They say a web page can drop such a file into Downloads without a click.

Harm: No harm foundControl level 3, Exceeded permissions

Not counted · can set the control floorEffect on the index

Near miss. Harm was plausible but prevented. It is rated only on what actually happened.

Sources

How we know

2 sources · single source. Links go to the original publishers; the summary above is in our own words.

  1. researchMarkdown Is the New Executable: Attackers Are Dropping Agent Instructions Into Your Downloads Folder From the WebForever · Oct. 22, 2026forever.security/blog/hijacking-ai-agents-via-downloads-folder/
  2. newsAI coding agents hijacked using AGENTS.md file in Downloads folderBleepingComputer · Oct. 25, 2026bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-fil…

Why this rating

No harm found; control failure level 3

Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.

Observed harm

No harm found

The malicious install was blocked on the customer's machine. The researchers ran the real malware only once, inside an isolated virtual machine.

Scope: The malicious package was not run.

The harm scale
  1. 1 Negligible Inconvenience, easily remedied.
  2. 2 Minor Limited, recoverable harm.
  3. 3 Moderate Material harm needing significant effort to remedy.
  4. 4 Severe Severe harm to health, rights, property or essential services.
  5. 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.

Control assessment

Exceeded permissions

Followed instructions planted in a file in its working folder and began installing a malicious package. With the same file in the researchers' isolated machine, it edited its machine-wide settings and started a second session with approvals and its sandbox turned off, without the user's agreement.

Reported beside the index. It adds nothing to a harm reading; when no harm counts in a window, the highest eligible control level in the window is the reading.

The control scale
  1. 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
  2. 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
  3. 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
  4. 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
  5. 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.

Rating rationale

On the customer's machine, Codex obeyed instructions planted in a file it reads on startup and began installing a known malicious package. The security firm watching that machine blocked the install before any code ran, so no harm occurred. When the researchers ran the same file in an isolated machine, Codex edited its own machine-wide settings and started a second session with approvals and its sandbox turned off. Those actions went beyond its task and were taken without the user's agreement. It exceeded its permissions on the user's own machine without reaching outside systems, which is control level 3.

The scales

Effect on the index

It does not count toward the Oct. 26 reading

The reading for the week to Oct. 26, 2026, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.

Not counted: no harm found. 21 other records behind the reading for that week.

The arithmetic
The reading for the week of Oct. 26, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceNo—
Worst documented harm, ksets the band1 Negligible1 Negligible
Harms at that level, nposition in the band33
Highest control level breachedsets the reading only when no harm counts— (none breached)— (none breached)
Readingrounded down15 Negligible harm15 Negligible harm
Inside the window of 4 weekly readings
Week toReadingBand
Oct. 26, 202615Negligible harm
Nov. 2, 20269Negligible harm
Nov. 9, 20269Negligible harm
Nov. 16, 20264Control failures only

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

2 logged. Every change to a rating is logged here, with the reason.

  1. v2
    Oct. 29, 2026

    Ratings confirmed by the editor. Changed: harm assessment.

  2. v1
    Oct. 29, 2026

    Added.

Cite and share

Use this record

Citation

Paperclip Index. “Planted AGENTS.md file in a user's Downloads folder hijacked OpenAI Codex into trying to install malware.” Record PI-0106. Reported Oct. 22, 2026; updated Oct. 29, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0106

Follow the record.

New incidents and signals, in your feed reader.