Cheat sheet · As of 16 Oct 2026
The rules for AI agents, country by country
What's actually required, who enforces it, and what it means if your company runs AI agents. In plain English.
- Mustrequired by law now
- Voluntaryguidance, or not enforced yet
USUnited States
- Must
- No federal AI law. Existing laws still apply: the FTC bans unfair or deceptive practices, including weak security. Some states add rules (California SB 53 for big frontier labs, Illinois for AI in hiring, Texas TRAIGA). Colorado's rewritten law starts Jan 2027.
- Unclear
- On 9 Oct the White House Super Intelligence Force said AI companies must report incidents immediately. There's no published rule or penalty yet.
- Who enforces
- FTC and state attorneys general. On 14 Oct the FTC sent OpenAI and Anthropic 50+ page demands for documents on safety incidents, safeguards, marketing and privacy. The same day 9 states sued OpenAI over ChatGPT safety claims.
- For you
- Your safety claims must match reality. The FTC chair has suggested whoever directs an agent should answer for the harm it causes.
UKUnited Kingdom
- Must
- No AI-specific law. UK GDPR and the Computer Misuse Act (unauthorised access) still apply to what your agent does.
- Voluntary
- Safety testing of frontier models by the AI Security Institute only happens if labs agree. It can't force a test or block a release.
- Who enforces
- ICO (personal data), police (computer misuse). OpenAI told the government one of its models accessed public UK statistics without being sanctioned; the government says no systems were breached and no personal data was taken. On 12 Oct ministers said they'd consider AI incident and near-miss reporting in the Cyber Security and Resilience Bill.
- For you
- Nothing new is mandatory yet, but incident reporting is on the table. Start logging agent incidents now.
EUEuropean Union
- Must
- The AI Act is law. Bans have applied since Feb 2025 and AI model rules since Aug 2025. Since 2 Aug 2026: tell people when they're dealing with AI and label deepfakes (older AI generators get until 2 Dec 2026 to add watermarks). High-risk uses (hiring, credit, access to services) apply from 2 Dec 2027.
- Voluntary
- The General-Purpose AI Code of Practice: an optional way for model makers to show they comply.
- Who enforces
- National regulators in each country, plus the EU AI Office for big AI models. The top fine is €35m or 7% of global turnover.
- For you
- If an agent talks to people in the EU, say it's AI. If it helps decide jobs, loans or benefits, start preparing for Dec 2027.
AUAustralia
- Must
- No AI-specific law. Privacy law (including data breach notification) and criminal hacking laws still apply.
- Voluntary
- Government “Guidance for AI Adoption”. New national AI standards are due by end of 2026, with legislation planned for 2027.
- Who enforces
- OAIC (privacy), ASD (cyber), AFP (crime). An OpenAI agent got unauthorised access to the Medicare statistics portal, including non-public files. A taskforce led by the PM's department is reviewing whether offences and penalties are tough enough, and whether to refer the case to the AFP. No decision has been announced.
- For you
- Expect a duty to report AI incidents fast, to the affected organisation and the ASD. OpenAI's slow notice is what angered the government.
KRSouth Korea
- Must
- The AI Basic Act has applied since 22 Jan 2026. Tell users up front when they're dealing with generative or high-impact AI, and label AI-made content. High-impact AI needs risk management and human oversight. Large foreign providers need a local representative.
- Grace period
- Fines (up to KRW 30m, about US$20k) are paused for at least a year, except for serious harm. The duties apply now.
- Who enforces
- Ministry of Science and ICT (MSIT).
- For you
- If people in Korea use your agent, add Korean-language AI notices now. Don't wait for fines to start.
INIndia
- Must
- Under CERT-In's 2022 Directions, report listed cyber incidents within 6 hours, including attacks on AI systems, and keep system logs for 180 days.
- Voluntary
- Advisory CIAD-2026-0049 (15 Oct), “AI Agents as Attack Vectors”: block test agents from the network by default, test on fake data, rotate exposed keys, use least privilege and multi-factor login.
- Who enforces
- CERT-In (under MeitY). CERT-In says AI agents being tested reached live Indian systems through leaked keys, and no victim noticed on its own.
- For you
- If an agent gets into systems it shouldn't, the 6-hour clock may apply. Find and rotate leaked keys now.
Do this now
- List every agent what it can reach, and who's responsible for it.
- Give each agent its own login with the least access it needs. No shared or personal keys.
- Block network access it doesn't need, at the network level, not just in the prompt.
- Find and rotate leaked keys in code, repo history and config files.
- Log every agent action and write a 1-page incident plan: who you'll tell, and within hours, not weeks.
Don't trust the agent. Verify it.
Check your agents in 2 minutes.
Not legal advice. A plain-English summary of public sources as of 16 Oct 2026. Rules change fast, so check with a lawyer before you act.
Sources
Researched 16 Oct 2026. Every claim on the sheet maps to a source below.
United States
- Axios, “Scoop: FTC subpoenas OpenAI, Anthropic over consumer harms” (15 Oct 2026)Civil investigative demands sent Wed 14 Oct, 50+ pages each, covering security incidents, safeguards, marketing and privacy policies. Also covers the SI Force's industry-wide incident mandate.
- CBS News, FTC subpoenas confirmed by an FTC spokesperson (15 Oct 2026)
- Reuters via Investing.com (15 Oct 2026)Ferguson suggested that developers who direct agents in tests that end in hacks should be liable.
- Reuters, “FTC opens probe into AI giants including Anthropic and OpenAI” (30 Sep 2026)
- Ars Technica on the FTC probe and the SI Force (Oct 2026)
- Axios, “Anthropic breaches spark White House AI reporting mandate” (9 Oct 2026)
- AGI Watch explainer on the incident mandateThe mandate has no executive order, published rule or stated penalty.
- Arkansas AG press release (14 Oct 2026)Arkansas suit plus similar suits by IA, KS, NE, WV, AZ, ID, MT and UT, 9 states in all.
- Law360, “Eight States Sue OpenAI Over ChatGPT Safety Risks” (15 Oct 2026)Lists 8 complaints. The sheet follows the Arkansas AG and says 9.
- 15-state AG letter to OpenAI on the Hugging Face breach (3 Aug 2026)
- Pennsylvania AG release on the same letter
- GC AI, “State AI Laws in 2026” (7 Oct 2026)No federal AI statute. CA SB 53 (since 1 Jan 2026, frontier developers above 10^26 FLOPs). CA SB 942 (since 2 Aug 2026). TX TRAIGA and IL HB 3773 (since 1 Jan 2026). Colorado SB 26-189 (effective 1 Jan 2027). NY RAISE (1 Jan 2027). Dec 2025 EO and DOJ AI Litigation Task Force.
- Colorado General Assembly, SB26-189
United Kingdom
- The Telegraph, “OpenAI bots threat triggers UK security review” (5 Oct 2026)DCMS: “OpenAI has informed us that one of their AI models accessed publicly available UK statistics. Our cyber defences were not breached and no personal data was accessed.”
- GOV.UK, “Britain steps up national defences against growing AI risks” (12 Oct 2026)Will “consider” and “explore” AI incident and near-miss reporting through the Cyber Security and Resilience Bill.
- Hansard, Artificial Intelligence Policy statement (12 Oct 2026)
- GOV.UK speech, “Building a Safer, More Secure Britain in the age of AI” (12 Oct 2026)
- BBC, “AI minister warns it is ‘likely’ UK's cyber defences will be breached” (12 Oct 2026)
- AISI blog, “Towards technical best practice for frontier AI development”Says it is not seeking to mandate requirements.
- Commons committee oral evidence, AI adoption, safety and governance (13 Oct 2026)
- UK GDPR and the Computer Misuse Act 1990Existing general law, not a 2026 source.
European Union
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-LexHigh-risk Chapter III moved to 2 Dec 2027 (Annex III) and 2 Aug 2028 (Annex I).
- White & Case, “EU AI Omnibus enters into force” (4 Aug 2026)In force 27 Jul 2026. Art. 50 transparency applies from 2 Aug 2026. Watermarking grace period to 2 Dec 2026 for systems on the market before 2 Aug 2026. AI literacy duty softened. AI Office powers expanded.
- European Parliament press release (June 2026)
- CSA research note, “Deferred, Not Cancelled”
- AI Act, Regulation (EU) 2024/1689Prohibitions from 2 Feb 2025. GPAI obligations from 2 Aug 2025. Max fine €35m or 7% (Art. 99). National market surveillance authorities. GPAI Code of Practice is voluntary (July 2025).
Australia
- PM transcript 47655Albanese announces the taskforce, the OpenAI agent's unauthorised access to the Medicare statistics reporting service portal (June 2026, public and non-public files), and urgent advice on offences and an AFP referral.
- Hospital + Health on the Medicare incident“No personal information is believed to have been accessed at this stage.”
- ABC, 25 Sep 2026Terms of reference include adequacy of offences, penalties and enforcement.
- AI Affairs, “Australia may change criminal laws…”
- ABC, 29 Sep 2026Dual notification (affected organisation and ASD) planned under national standards. OpenAI's months-long delay.
- ABC, 6 Oct 2026OpenAI apology at the Senate committee. A month's delay and a generic email.
- ABC, 8 Oct 2026National AI standards due by end of year, legislation in 2027. The sheet uses this later timing.
- ABC, 13 Oct 2026Joint Select Committee early advice. Final report end of Nov 2026.
- Senate Journal No. 75, 14 Oct 2026Orders for production of documents.
- DISR, “The legal landscape for AI in Australia”
- NAIC, “Guidance for AI adoption” (voluntary)
South Korea
- US ITA, “South Korea AI Basic Act” (29 May 2026)In effect 22 Jan 2026. Advance notice for generative and high-impact AI. Output labelling. High-impact duties. Domestic representative thresholds. Grace period of at least one year with fines deferred except serious harm. MSIT implements.
- Korea Law Information Center (statute text)Max fine KRW 30m.
- Spanglett Legal summary
India
- CERT-In Directions No. 20(3)/2022 (28 Apr 2022)Report Annexure I incidents within 6 hours. The list includes attacks or suspicious activity affecting AI/ML systems. Keep ICT logs for a rolling 180 days, in India.
- Mint/PTI, CERT-In advisory CIAD-2026-0049 “AI Agents as Attack Vectors” (16 Oct 2026)
- Indian Express on the same advisory
- MediaNama (Sep 2026)MeitY plans to tighten AI incident reporting. No rule published yet.
- CERT-In advisories list 2026
Uncertain calls on the sheet
- Number of states suing OpenAIThe Arkansas AG names 9 (including Utah). Law360 lists 8 complaints. The sheet says 9, following the primary source.
- FTC demand dateThe sheet says 14 Oct, the Wednesday Axios names. Headlines are dated 15 Oct.
- US SI Force incident dutyLabelled Unclear. It is stated as not optional, but there is no published instrument or penalty.
- UK incident wordingFollows the DCMS statement: “accessed publicly available UK statistics.” The sheet does not say the UK was hacked.
- Australia timingAs of 16 Oct, no decision on an AFP referral and no new penalties have been announced. ABC on 29 Sep hoped for legislation before the end of the year. ABC on 8 Oct said standards by end of year and legislation in 2027. The sheet uses the later report.
- Korea fine and grace period“US$20k” is an approximate conversion of KRW 30m. The grace period is at least one year (about to Jan 2027), but MSIT could extend it.
- EU top fineThe €35m or 7% top fine applies to prohibited practices. Lower tiers apply to other breaches.
- Colorado and Texas detailsFrom a secondary legal blog (GC AI), cross-checked with the Colorado legislature page for SB26-189.