הניתוח

ניתוחים

פוסטים שבועיים על בטיחות Physical AI. מגובה בנתונים. מצוטט ממקורות.

פרסום מקצועי לאנשים שבונים, מסדירים, ומממנים את העשור הבא של Physical AI. כל פוסט מתחיל ממקור ראשוני, מצטט את המקור, וטוען טיעון אחד בכל פעם. ללא ספקולציות היכן שיש נתונים. ללא נתונים היכן שספקולציה תהיה כנה יותר.

מומלץ

After the attack: the security patch. When does it make you a new manufacturer?

רגולציה2026-10-292 דקות קריאה

Part 9 of 12. After an attack, the robot gets patched, and sometimes its model gets retrained. Under EU machinery law, a modification 'by physical or digital means' that wasn't foreseen, raises a risk and needs new protective measures can turn whoever made it into the manufacturer. The Cyber Resilience Act wants security fixes shipped apart from new features where feasible.

קראו את הפוסט המומלץ

Illustration: AI-generated

רגולציה

After the attack: the security patch. When does it make you a new manufacturer?

29 באוק׳ 2026קרא →

רגולציה

After the attack: Europe's clock starts at 24 hours. The US clock depends on the sector.

Part 8 of 12. Since 11 September 2026, the Cyber Resilience Act gives makers of most connected products, robots included, 24 hours to warn about an actively exploited vulnerability. In the US, vehicle makers report safety defects within 5 working days, and the federal critical-infrastructure reporting rule still isn't in force.

28 באוק׳ 2026קרא →

הנדסה

During the attack: who reaches the safe state?

Part 7 of 12. EU machinery law says the stop must beat the start, and the emergency stop must work 'at all times, regardless of the operating mode'. It doesn't say where the stop must run. When an attacker holds the AI's computer, that question decides whether the robot stops.

27 באוק׳ 2026קרא →

רגולציה

Before the attack: the US has no physical AI cybersecurity law. It has sector rules.

Part 6 of 12. Europe writes one horizontal duty for machines from 20 January 2027. The US splits it: a binding cybersecurity duty for medical devices since March 2023, recall rules and voluntary guidance for vehicles, no OSHA standard for robots, and a voluntary label for consumer devices.

22 באוק׳ 2026קרא →

תקנים

Before the attack: five documents touch physical AI cybersecurity. One mentions AI, and it sets no requirements.

Part 5 of 12. IEC 62443, IEC TS 63074, IEC 61508, ISO/IEC TR 5469 and the draft FprEN 50742 all touch the security side of machine safety. Read through their publishers' own public pages, only one names AI, and it's a Technical Report. None is cited yet under the EU Machinery Regulation.

21 באוק׳ 2026קרא →

רגולציה

Before the attack: four duties EU law puts on a physical AI maker from 20 January 2027

Part 4 of 12. The Machinery Regulation answers 'what must the maker build in?' with four lines, all in the safety annex: withstand the attacker, protect the software and data the safety depends on, keep a self-evolving machine inside its task, and send machine-learning safety components to a third party. The EU hasn't cited a single harmonised standard under the Regulation yet.

20 באוק׳ 2026קרא →

הנדסה

If the attacker owns the AI's computer, what else does he own?

An attacker who owns an AI's computer owns every safety check running on it. My own papers name the sharing and put an estimate on it: across anonymised architectures, software-only redundancy approaches 100% common-cause failure. A nuclear regulator's 1994 report doubted the same assumption decades earlier. This post asks who reaches the safe state. It does not answer.

15 באוק׳ 2026קרא →

הנדסה

They talked the physical AI into it

RoboPAIR jailbroke three LLM-controlled systems, hitting 35 for 35 on a real quadruped, a real wheeled robot, and a set of driving video clips with no vehicle at all. The paper calls only the quadruped's actions physically realized. Its own limit: none of the three reason about harm on their own.

14 באוק׳ 2026קרא →

הנדסה

A hacked chatbot says something wrong. A hacked physical AI moves.

Part 1 of 12 on Physical AI cybersecurity as a safety problem. When an AI drives a motor, a compromise ends in motion. From 20 January 2027, EU machinery law makes withstanding attackers a safety duty. The US splits the question by sector.

13 באוק׳ 2026קרא →