The Index
The Paperclip Index
Documented harm from AI, read daily.
Control C4 Level 4, reached outside systems1 record in the last 30 days
No harm counting today. The reading is the highest control level breached, 4: a level, not a count (PI-0110). 2 records reported in the last 30 days.
Records behind this reading · 1
- PI-0110 · AI hacking agent in a researcher's benchmark test escaped its sandbox and pulled customer records from WorkOSControl level 4, reached outside systems: sets the reading · extent undisclosed
Only these records move the reading. How it is calculated
Status watch · 13
Current assessments across all report dates; leaving the window does not mean resolved.
- PI-0067 · OpenAI research agents posted user-provided images to image-hosting sites in 53 instancesOngoing · Assessed Sept. 25, 2026
- PI-0077 · Early Claude Opus 5.5 snapshot wrote a command to send secrets to an external host during internal useOngoing · Assessed Sept. 22, 2026
- PI-0084 · Pennsylvania sues Character.AI, alleging chatbots posed as licensed psychiatrists and gave a fake license numberOngoing · Assessed May 5, 2026
- PI-0105 · OpenAI model asked to read a public dashboard broke into its owner's internal systems and changed their settingsReached outside systems, outcome not reported · Assessed Oct. 15, 2026
- PI-0092 · Claude Mythos Preview exploited an injection flaw to run commands on a university server during an evaluationReached outside systems, outcome not reported · Assessed Oct. 9, 2026
- PI-0083 · OpenAI agent used crafted queries to infer hidden database details from a NSW parks agency's fire history serviceReached outside systems, outcome not reported · Assessed Oct. 3, 2026
- PI-0064 · OpenAI research agent accessed non-public files on an Australian government Medicare statistics portalReached outside systems, outcome not reported · Assessed Oct. 6, 2026
- PI-0073 · Gemini guessed a password and used leaked credentials to get into three real companies during a hacking testReached outside systems, outcome not reported · Assessed Sept. 19, 2026
- PI-0074 · Early Claude Opus 4.6 broke into a real outside machine during a hacking test and read one person's dataReached outside systems, outcome not reported · Assessed Sept. 9, 2026
- PI-0072 · Meta's Muse Spark 1.1 broke into an outside company's service after a testing vendor left it onlineReached outside systems, outcome not reported · Assessed Aug. 14, 2026
- PI-0071 · Anthropic models in a hacking test reached three real companies, took credentials and published a malicious packageReached outside systems, outcome not reported · Assessed Sept. 9, 2026
- PI-0060 · Personal agent exploited a gym booking flaw to remove another member from a waitlistReached outside systems, outcome not reported · Assessed Aug. 10, 2026
- PI-0049 · Alibaba-linked ROME agent opened an outbound tunnel and diverted training GPUs to crypto miningReached outside systems, outcome not reported · Assessed March 12, 2026
- Reading
- 30-day average
Today
Hover, focus or tap a block to inspect its record.
The harms that count
Incidents by impact
NegligibleCatastrophic
The last 30 days
What was reported, and how control was lost
Change from the 30 days before
How control was lost
Recent disclosures
- Incident, harm counts. Darker means higher harm level.
- Incident, no harm that counts. Logged, not in the reading.
- Signal. Policy or capability news. Context only, never counted.
As a table
Over time
How control was lost, and how reports arrive
How control was lost, by quarter
When it happened, when it surfaced
Control lost. Harm done?
Inspect matrix records
All records shown in the matrix
Explore a mark to see the record behind it.
Selected reports, not a census.
The scale
One harm opens its band. Ten at that level fill it. The calculation