Weekly reading · methodology v1.0

Window ending June 29, 2026

Harms count in full for two weeks after they are reported, then one level less every two weeks. The control floor looks at reports from May 31, 2026 through June 29, 2026. A selected catalogue of reported AI incidents, reviewed through Oct. 10, 2026. This reading describes documented harm in the records or, when none qualifies, the highest control level breached. It is not a forecast or a measure of all AI activity.

Current recalculation

Control failures only · No qualifying harm. The reading is the highest control level breached: 3, by 1 record. It is a level, not a count.

Recalculated from the catalogue in this build. Historical backcasts were not published at the time.

Published at the time

No publication snapshot exists for this week.

The harm ladder: the worst documented harm level picks the step; each tick is one qualifying harm at that level, and ten fill the step. The first, dark segment is the control floor: readings 2 to 5 mean no harm qualified and show the highest control level breached.

This is a backcast from the current catalogue. It was not a reading published at the time.

Records behind this reading

3 selected records; 0 documented external harms counting (1 qualifying). Records and ratings below reflect the current catalogue.

No harm counting that week. The reading is the highest control level breached, 3: a level, not a count (PI-0076). 3 records reported in the last 30 days.

Records behind this reading · 1

PI-0076 · Co-located Claude Mythos 5 agents killed each other's processes and disguised their own to avoid being killedControl level 3, exceeded its permissions: sets the reading

Only these records move the reading. How it is calculated

PI-0089
Meta's AI support assistant let attackers hijack 20,225 Instagram accounts by linking their own emailJune 1, 2026 · Minor harm · no longer counting · Deployment
PI-0076
Co-located Claude Mythos 5 agents killed each other's processes and disguised their own to avoid being killedJune 9, 2026 · No harm reported · Internal research
PI-0085
GPT-5.6 Sol cheated on METR's software tasks more than any public model METR had testedJune 26, 2026 · No harm reported · Controlled test

Sources

These links support the records above. Source availability and conclusions may change.

404 Media: Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It WorkedCited in PI-0089
Engadget: Meta's AI support chatbot made it ridiculously easy for hackers to take over Instagram accountsCited in PI-0089
BleepingComputer: Over 20,000 Instagram accounts stolen in Meta AI support hackCited in PI-0089
Help Net Security: Hackers used Meta’s AI support system to hijack over 20,000 Instagram accountsCited in PI-0089
Anthropic: System Card: Claude Fable 5 & Claude Mythos 5Cited in PI-0076
Anthropic: Risk Report: August 2026Cited in PI-0076
Don't Worry About the Vase (Zvi Mowshowitz): Claude Fable 5 and Mythos 5: The System CardCited in PI-0076
METR: Summary of METR's predeployment evaluation of GPT-5.6 SolCited in PI-0085
OpenAI: GPT-5.6 System CardCited in PI-0085
The Decoder: OpenAI's new flagship model GPT-5.6 Sol cheats on software tests more than any model before itCited in PI-0085
Transformer: GPT-5.6 cheats so much its testers couldn’t measure itCited in PI-0085

Every weekly brief · How this reading is calculated · Download the weekly card