From the journal

AI penetration testing: authorisation, third-party systems and liability

An AI penetration-testing service should start only when the provider can connect each target, permitted act and foreseeable effect to a valid authority. A customer’s domain list, cloud subscription and test credentials are useful evidence. Verify the entitled party’s grant for each repository, host, integration and neighbouring tenant within scope. An indemnity may allocate recoverable loss between the parties; third-party rights require authority from the person entitled to grant them.

23 min read

The practical decision is asset specific. Accept work where authority, platform conditions, data arrangements and operational controls are established. Narrow a test to the customer-controlled layer where wider rights are missing. Defer a target while material permissions or local requirements remain unresolved. Refuse activity that requires knowingly crossing an unapproved boundary or continuing after permission is withdrawn.

AI penetration testing legal requirements vary across markets. Some access offences distinguish prohibited purpose from access to a forbidden area. Others make the terms of permission, particular harmful effects or additional intent central. A licensing requirement, research defence, disclosure programme and contractual safe harbour each answer different questions. The following comparison brings those distinctions into the authorisation and contract decisions a provider must make.

Establish authority over the asset and the act

Start with a written authority map. Identify the legal entity controlling each relevant layer, the person approving it, the source of that person’s powers and any restriction on delegation. An affiliate’s systems should have their own entry. A corporate relationship or a person’s administrative access should not be treated as proof that the customer can commission all proposed activity.

The UK Computer Misuse Act makes the relevant kind of access and consent from a person entitled to control it central. Singapore also distinguishes access and modification and the person entitled to control them. These rules support an authority record that is more precise than “permission to test the network”. UK CMA, s 17(5), (8); Singapore CMA, s 2(5), (7)–(8).

  • Customer application and accounts. Evidence to obtain before acceptance: Exact application and account identifiers, the customer’s control rights, signatory powers and the permitted use of each account.
  • Affiliate or separately owned repository. Evidence to obtain before acceptance: The relevant entity’s own grant or an evidenced delegation covering this testing and onward appointment of the provider.
  • Cloud workload. Evidence to obtain before acceptance: Customer authority over the workload, the applicable hosting agreement and platform testing conditions, plus any required special approval.
  • Shared infrastructure or another tenant. Evidence to obtain before acceptance: A separate entitled-party grant covering that layer and its effects, or an express exclusion with workable separation.
  • Personal or confidential information. Evidence to obtain before acceptance: Processing instructions, lawful purpose, permitted recipients and transfers, evidence limits and retention rules, assessed separately from system access.

The list above is a proposed acceptance standard. Not every legal system requires a signed document for every form of consent. Brazil’s computer-invasion offence, for example, refers to express or tacit authorisation and requires a specified purpose. Written evidence nevertheless helps establish what was permitted, who granted it and when it ended. Brazil Penal Code, art 154-A.

Divide permission into reading, copying, executing, modifying and producing effects. Authority to inspect a test account does not necessarily include extracting unrelated records, modifying production data or interrupting a shared service. Define the approved environment, identifiers, time window, accounts, data categories and excluded effects. Obtain scope approval before using a newly discovered credential or extending work to a newly discovered dependency.

Withdrawal must be operationally effective. Identify who can revoke permission, how the provider receives that instruction and how it stops queued and active work. Preserve the approved scope version and the time a withdrawal became effective. The contract should require a new documented decision before restart. The wider separation between a model’s proposal, an authorised approval and execution is addressed in Licentium’s Model Context Protocol (MCP) permissions analysis.

In the United States, Van Buren v United States held that a person entitled to obtain information through a computer does not “exceed authorised access” under the Computer Fraud and Abuse Act (CFAA) merely by obtaining it for an improper purpose. The Court distinguished access to information in areas that are off limits. It expressly left open whether the boundary must be technological or may also arise from contracts or policies. The decision therefore does not convert a working credential into unrestricted testing authority. 593 US 374 (2021), slip opinion pp 12–15 and n 8.

The offence and remedy still matter. CFAA provisions separately address obtaining information, fraud and damage. Its civil remedy requires the statutory damage or loss conditions; it is not available for every policy breach. Section 1030(g) also excludes an action under that subsection for negligent software design or manufacture. That exclusion does not erase other causes of action or protect conduct satisfying a different applicable rule. California Penal Code section 502 separately addresses knowing, unpermitted access, use, copying and interference, with its own civil remedies. 18 USC 1030(a), (e), (g); California Penal Code, s 502(c), (e).

Canada requires its own analysis. Criminal Code section 342.1 addresses conduct undertaken fraudulently and without colour of right, including obtaining computer services and certain interception or password activity. Section 430 separately addresses wilful damage and interference with computer data or lawful use. The provider should document actual authority and the basis for relying on it, then assess the particular conduct and mental element. An accidental deviation should not be labelled a completed crime without that analysis. Criminal Code, ss 342.1 and 430; s 430 operative text.

Australia and New Zealand make another useful distinction. Australian federal law says access is not unauthorised merely because the person has an ulterior purpose; New South Wales has its own comparable qualification. New Zealand section 252(2) excludes the access offence where a person is authorised to access the system, even for a purpose different from that authorised. These provisions do not remove separate rules about restricted data, dishonest acquisition, modification or impairment. A scope schedule must address those acts and effects individually. Australian Criminal Code, ss 476.2, 477.2–477.3, 478.1; NSW Crimes Act, ss 308B, 308H; NZ Crimes Act, ss 249–254.

For UK testing, assess the intended access and knowledge under section 1 separately from impairment under section 3 and serious-damage conduct under section 3ZA. The Crown Prosecution Service (CPS) guidance also distinguishes permitted access used for an improper purpose from access to material the person was not entitled to obtain. Neither a blanket “all purpose breaches are hacking” rule nor a blanket “credentials make it lawful” rule is accurate. CMA, ss 1, 3, 3ZA, 17; CPS guidance, relevant case-law discussion.

Research protections must match the proposed activity

The US Department of Justice’s CFAA charging policy directs that good-faith security research should not be charged. Its conditions concern a security purpose, avoidance of harm and use of the information primarily to promote security. Extortion falls outside that policy. It is an internal federal prosecution policy, not an enforceable immunity from private proceedings, state law or other offences. A commercial provider should not substitute it for its customer and target-authority records. Justice Manual 9-48.000, B and C.

France illustrates the danger of extending a narrow provision. Penal Code article 323-3-1 recognises legitimate purposes, including security research, in the provision concerning tools and similar means for computer offences. It does not supply a general defence to fraudulent access, interference or data extraction under articles 323-1 to 323-3. Defence Code article L2321-4 gives a specific reporting protection for qualifying good-faith vulnerability reports made exclusively to the National Cybersecurity Agency of France (ANSSI); it does not authorise every act used to discover the vulnerability. Penal Code, arts 323-1–323-3-1; Defence Code, art L2321-4.

EU Directive 2013/40 likewise distinguishes illegal access, interference, interception and tools. Its minimum access rule includes intentional conduct, lack of right and infringement of a security measure, at least for non-minor cases. Member States supply the operative national offences and remedies. A provider testing French systems must analyse French law, rather than treating the Directive as a single EU testing licence. Directive 2013/40/EU, arts 2–7, 10–12.

Chile’s Law 21.459 article 2 excludes criminal punishment for its first-paragraph access offence only when the system’s responsible person is domiciled in Chile and every listed condition is met. The researcher must be on the National Cybersecurity Agency (ANCI) register, notify ANCI before access, report the access and vulnerabilities to ANCI and the responsible person as soon as the access has occurred, and meet ANCI’s responsible-disclosure conditions. Testing must stay within what is necessary to establish a vulnerability and avoid fraudulent purposes or an intent to acquire or use the system’s information. Excluded methods include denial of service, physical testing, malicious code, social engineering, and alteration, deletion, exfiltration or destruction of data. Public disclosure of vulnerability information is also excluded. Systems outside the State Administration require the responsible person’s consent. The former article 16 was repealed. Evidence actual registration and prior notice before using this route. Law 21.459, art 2, as amended by Law 21.663.

Kenya’s testing exception is also limited: section 18(3)(a) concerns the tools provision, not unrestricted access under section 14. Section 14(2) joins its entitlement and consent tests with “or”, creating a material issue for an outside tester relying only on consent. Obtain a grant covering relevant delegated control as well as consent; where that cannot be established, resolve the statutory interpretation before activation. Critical-information-infrastructure audits have a further statutory authorisation route. Computer Misuse and Cybercrimes Act, ss 13(6), 14, 18.

Hong Kong’s proposed cyber-dependent-crime reforms must also remain separate from enacted law. As at 18 November 2026, the Law Reform Commission’s research-defence recommendations are proposals. In Secretary for Justice v Cheng Ka Yee, the Court of Final Appeal’s conclusion about section 161 of the Crimes Ordinance concerned use of one’s own computer without accessing another computer. It was not a general security-testing exemption. 2019 HKCFA 9, paras 44–48; LRC report and recommendations.

Provider status and customer sector can add conditions

Singapore regulates penetration-testing service providers under its Cybersecurity Act. The service, delivery entity and commercial structure must be classified before relying on a licence or an exemption. Official guidance includes overseas providers and resellers in relevant circumstances, while distinguishing a software-only supply. The customer’s permission and the provider’s licence perform different functions. Section 31 can also affect recovery of fees for unlicensed provision. Cybersecurity Act, ss 24, 31 and Second Schedule; CSRO FAQ.

The licensing terms affect delivery records and data use. The licensee must retain the prescribed engagement records for at least three years after the occasion to which each record relates. Define separate lawful retention periods for copied target records. The licence conditions address customer-data use and maintenance of active Cyber Trust Mark Promoter (Tier 3) or equivalent certification. As at 18 November 2026, the Cybersecurity Services Regulation Office (CSRO) FAQ gives applicants and licensees a grace period until 31 December 2026. Check the condition applicable to the actual licence and delivery date. Cybersecurity Act, s 29; 2022 service-provider regulations, reg 4; CSRO conditions of licence, cl 6.1; CSRO FAQ, Q9, Q23, Q36–37.

In Saudi Arabia, verify the applicable National Cybersecurity Authority (NCA) service-provider registration and classification. A consultation on a new licensing framework is not evidence that its proposed licensing conditions have commenced. The Anti-Cyber Crime Law also requires offence-specific analysis: its definition of unlawful access should not be read as a standalone offence covering every contractual deviation. Identify the actual provision, conduct, purpose and territorial connection. NCA registration materials; RFCS consultation; Anti-Cyber Crime Law, arts 1, 3–6.

The United Arab Emirates (UAE) requires identification of the relevant federal, emirate, financial-zone and sector rules. The federal cybercrime definition includes access without permission or contrary to its terms. Financial-zone status does not remove federal criminal law. Dubai government-related work may involve a separate provider qualification, while banking work requires its own regulatory assessment. A Dubai International Financial Centre (DIFC) or Abu Dhabi Global Market (ADGM) incorporation certificate therefore cannot settle the entire engagement. Federal Decree-Law 34/2021, art 1 and relevant offences; Federal Law 8/2004, art 3; Dubai certified-provider framework.

For selected EU financial entities, threat-led penetration testing (TLPT) under the Digital Operational Resilience Act (DORA) adds a different layer. Articles 26–27 address validated scope, production systems, third-party participation, qualified testers and insurance. The detailed 2025 rules require control arrangements, risk management, evidence and restoration measures. They provide for managing shared-service risks, including a pooled-testing route in specified circumstances. Ordinary automated vulnerability testing does not become compliant TLPT simply because the customer is a fintech. DORA, arts 26–27; Delegated Regulation 2025/1190, arts 3–8, 9(11)–(12), 13, 16(4)–(5).

Hong Kong’s critical-infrastructure regime similarly depends on operator and system designation. Its code addresses qualified testing and alternatives where operational technology makes ordinary testing unsuitable. The provider should obtain the customer’s designation and system scope, rather than assuming all businesses have identical obligations. OCCICS Code of Practice, security assessment and testing requirements.

A cloud account does not authorise every hosting layer

Platform policies supply permissions and conditions within their own scope. AWS permits specified security assessments of listed customer services without prior approval, but excludes particular activities and requires prior approval for all command-and-control testing. A request submitted in advance is not evidence that approval was granted. AWS penetration-testing policy.

Microsoft generally does not require prior approval for testing permitted customer resources, but its rules require an agent acting for a customer to obtain the resource owner’s explicit written authorisation. Google Cloud permits testing of a customer’s own projects subject to its terms and acceptable-use policy. As at 18 November 2026, that policy expressly addresses agentic AI; permission does not extend to disrupting other customers. Microsoft testing rules; Azure testing guidance; Google Cloud FAQ; Google Cloud AUP.

GitHub’s programme concerns its defined targets and rules. Its conditional legal safe harbour cannot bind a third party. Permission to examine a customer repository does not by itself authorise testing GitHub’s infrastructure or another repository owner’s deployment. GitHub programme rules; legal safe harbour.

Save the applicable policy version, approved services, exclusions and any written response with the engagement record. Recheck them before launch and when the target, method or platform service changes. Use this proposed evidence-management control alongside the required grants and approvals.

Allocate responsibility for what the agent actually does

The California Attorney General’s 16 November 2026 complaint against OpenAI concerns alleged conduct during model development and testing, including alleged third-party harm and inadequate safeguards. Its pleaded claims are under the Unfair Competition Law (UCL) and public-nuisance law. The pleading alleges those matters and requests controls and remedies. The commercial provider’s actual conduct and each applicable legal test still require separate assessment. Official complaint, paras 95–119.

The underlying statutes have distinct elements and remedies. California’s UCL addresses unlawful, unfair or fraudulent business practices. The Attorney General’s public-enforcement route under section 17204 differs from private standing, which requires injury in fact and lost money or property resulting from the practice. Injunctions, restitution and civil penalties have separate statutory conditions. Under Civil Code section 3480, public nuisance affects an entire community or neighbourhood, or a considerable number of persons, at the same time. Section 3494 permits abatement by a public body or officer authorised by law. Use these pleaded theories to assess containment evidence and representations in the California setting. Business and Professions Code, ss 17200, 17203–17206; Civil Code, s 3479, s 3480 and s 3494.

The Ninth Circuit’s August 2026 Amazon v Perplexity opinion answers a narrower attribution question. On the preliminary-injunction record describing Comet’s browser and server architecture, Amazon was unlikely to establish that Perplexity itself accessed Amazon’s computers. The court reserved different facts involving the provider’s control and liability in other contexts. It did not hold users liable, decide every CFAA element or immunise all agent developers. No 26-1444, pp 15–18 and n 4. Licentium’s browser-agent case analysis addresses that development separately.

For a penetration-testing service, record who selects targets, holds credentials, controls execution, changes instructions and can stop the agent. A provider operating server-side jobs or independently expanding targets should not assume that a decision about the architecture of a user’s browser resolves its exposure. Automation also does not remove the need to prove the fault, causation, damage and other elements of the applicable claim.

South Africa combines its Cybercrimes Act with the separate responsible-party and operator duties of the Protection of Personal Information Act (POPIA). POPIA requires appropriate safeguards and a written operator arrangement; its civil remedy in section 99 is not confined to proof of intentional or negligent conduct, although statutory defences remain. The customer contract should therefore allocate processing, supervision and cooperation obligations without assuming that ordinary negligence wording exhausts the exposure. Cybercrimes Act, ss 2–7; POPIA, ss 19–22, 99.

Brazil’s General Data Protection Law (LGPD) likewise separates controller and operator responsibility, including specified circumstances of joint liability and available defences. Mexico’s federal computer offences distinguish ordinary, state and financial-system situations, including particular conduct by an authorised user. An account holder’s consent is therefore only one part of the analysis. Mexico’s replacement 2025 private-data law must be used for security and notification duties. LGPD, arts 42–46; Mexican Federal Penal Code, arts 211 bis 1–7; 2025 private-data law, arts 18–20.

Location should be recorded at the operator, execution service, target, affected-person and effect levels. The CFAA can cover foreign computers affecting US commerce or communication; UK offences have specified significant-link rules; France’s electronic-network rule links the covered offences to France where they are attempted or committed against an individual resident there or a legal entity with its registered office there. Assess each statutory connection separately from the contract’s governing law. 18 USC 1030(e)(2); UK CMA, ss 4–5; French Penal Code, art 113-2-1.

Contract terms must fit mandatory local rules

The customer should warrant the authority it actually has and disclose known restrictions. The provider should commit to the agreed scope, professional care, approved personnel and effective suspension. Define deliverables and acceptance criteria without promising discovery of every vulnerability. Tie any authority indemnity to recoverable loss caused by a false representation or breach, with allocation for the provider’s own expansion or fault.

Liability caps need a local-law analysis. English-law business terms can engage the negligence and standard-term controls of the Unfair Contract Terms Act (UCTA) and its reasonableness test. Singapore’s UCTA has comparable questions, but its precise scope, exceptions and choice-of-law rule must be applied. In both systems, a governing-law choice alone can engage a specific limitation on the operation of the relevant provisions where the proper law would otherwise be foreign. UK UCTA, ss 2–3, 11, 27 and Schedule 1; Singapore UCTA, ss 2–4, 11, 27 and First Schedule.

In Quebec, Prelco upheld the negotiated allocation in the particular non-consumer, non-adhesion services contract before the Supreme Court. The judgment preserved mandatory restrictions, including those concerning intentional or gross fault and bodily or moral injury; it did not validate every Canadian liability cap. France separately treats as unwritten a clause depriving an essential obligation of its substance, while its contract-damage rule addresses foreseeability and gross or deliberate fault. 2021 SCC 39, paras 43–49, 84–99; French Civil Code, arts 1170 and 1231-3.

The UAE’s federal Civil Transactions Law took effect on 1 June 2026. Article 257 invalidates exemption from, or mitigation of, liability for a harmful act. Contractual non-performance and agreed compensation have their own provisions. Saudi Civil Transactions Law article 173 separately restricts contractual exemptions for fraud or gross fault and exclusions of harmful-act liability. Do not transplant one fee-based cap across these rules, DIFC law and ADGM law. UAE Federal Decree-Law 25/2025, arts 257, 334, 340; Saudi Civil Transactions Law, art 173.

California Civil Code section 1668 restricts contracting away specified wrongdoing. Section 2773 invalidates an indemnity for a future act where the actor knows it is unlawful when doing it. Those provisions reinforce the distinction between allocating insurable commercial risk and promising protection for knowingly unauthorised activity. They do not justify an unqualified assertion that every negligence limitation is void. Civil Code, ss 1668 and 2773.

Australia’s consumer definition can include a business service purchase costing AUD 100,000 or less, applying the statutory valuation rules. Covered service guarantees cannot simply be removed by a general cap. For the separate small-business unfair-terms regime, at least one business party must meet the statutory size test: fewer than 100 employees or turnover below AUD 10 million for the relevant last income year. Apply the counting rules and the separate standard-form and unfairness tests. Australian Consumer Law, ss 3, 23–24, 60–64A; Competition and Consumer Regulations 2010, reg 77A.

New Zealand requires a separate Consumer Guarantees Act scope check. A business buyer can qualify for services of a kind ordinarily acquired for personal, domestic or household use, subject to section 2’s exclusions. Where the Act applies, section 43 requires a written agreement, supply and acquisition in trade, all parties in trade and agreeing to exclusion, and a fair and reasonable term. NZ Consumer Guarantees Act, ss 2, 28–32, 43.

South Africa’s consumer-law restrictions also depend on the Consumer Protection Act’s actual application, including relevant juristic-person thresholds. Where applicable, risk notices and gross-negligence restrictions matter. Brazil and Mexico require attention to good faith, mandatory rights and their own restrictions on liability arrangements. The contractual annex should identify the actual local restriction and affected clause instead of inserting an unexplained worldwide carve-out. South African CPA, ss 5, 48–51; Brazil Civil Code, arts 421–422; Mexican Federal Civil Code, art 2106.

Insurance should be checked against the actual activity, territories and losses. Obtain the wording or a specific confirmation covering automated security testing, relevant subcontractors, data exposure and contractual liabilities. Record exclusions, deductibles and claims-notice conditions. A generic insurance certificate does not prove that an out-of-scope incident or an assumed indemnity is covered. For DORA TLPT, the statutory tester-insurance requirements remain a separate condition.

Build the scope schedule around evidence and decisions

The authorisation schedule should connect six records: the engagement and accountable people; the authority for each asset; approved operations and effects; data handling; change control; and incident responsibilities. Record the actual grant and corroborating evidence. Match the legal entity, account identifiers, delegated powers and current platform conditions. Resolve material inconsistencies before acceptance.

Specify the approved model and execution services, automation level, subcontractors and processing locations. Require fresh approval where a change affects targets, permitted effects, autonomy, recipients, retention or the ability to stop. A version change with no material effect can follow the agreed ordinary change process; a new recipient of target data should not be treated as a cosmetic update.

Data use needs its own basis. Under EU GDPR article 28, a processor acts on documented instructions, and subprocessor appointments require the specified prior written authorisation. A processor deciding its own purposes and means can become a controller for that processing. Security-testing duties under article 32 do not themselves authorise intrusion into somebody else’s system. New Zealand section 11 likewise distinguishes processing solely on behalf of another agency from own-purpose use. GDPR, arts 28, 32; NZ Privacy Act, s 11.

Use minimised or test data where adequate. Define permitted evidence, recipients, transfer arrangements and deletion dates. Do not include independent model training or product improvement within a routine testing instruction. Preserve necessary approvals, relevant action records and incident chronology without placing reusable credentials or unnecessary personal payloads in ordinary logs. Where deletion and preservation obligations conflict, identify the exact material and legal duty rather than imposing an unlimited hold.

The provider’s responsible human must be reachable during execution and able to stop it. Authority expiry, revocation, an unexpected tenant, unapproved effects, a platform instruction or loss of effective controls should suspend the affected work. Record detection and effective-stop times. Restart should require the missing authority or control to be restored, an approved scope version and the designated decisions. Customer silence is not a new grant.

Incident clocks require separate entries

The contract should require rapid internal escalation early enough for each party to meet its own duties. A short operational target is a negotiated term, not a replacement for a statutory trigger. Identify who assesses the event, who has each legal duty, the recipient, the awareness or assessment time and the proof of notification.

  • EU GDPR. Difference the incident schedule must preserve: A processor informs the controller without undue delay. The controller’s qualified notification duty is without undue delay and, where feasible, within 72 hours after awareness; affected-person notice has a separate high-risk test. Arts 33–34.
  • Singapore. Difference the incident schedule must preserve: Following a reasonable and expeditious assessment, notice to the Personal Data Protection Commission is as soon as practicable and no later than three calendar days after the day the breach is assessed as notifiable. An intermediary’s escalation duty is without undue delay. PDPA, ss 26C–26D.
  • Canada. Difference the incident schedule must preserve: Where the Personal Information Protection and Electronic Documents Act (PIPEDA) applies, the organisation reports to the Privacy Commissioner of Canada and notifies the individual if it is reasonable to believe the breach creates a real risk of significant harm. Both duties apply as soon as feasible after determining that the breach occurred. Establish federal and provincial routing first. PIPEDA, s 10.1.
  • South Africa. Difference the incident schedule must preserve: POPIA separates immediate operator escalation from the responsible party’s notification as soon as reasonably possible, subject to its conditions. The Cybercrimes Act’s section 54 reporting provision was excluded from the cited commencement proclamation. POPIA, ss 21–22; Proclamation R42/2021.
  • Brazil and Mexico. Difference the incident schedule must preserve: Brazil’s qualifying controller notices generally use three working days under the regulation of the National Data Protection Authority (ANPD). Mexico’s article 19 requires immediate affected-person notice for the specified significant-impact breach. ANPD Resolution 15/2024; Mexican private-data law, art 19.
  • Australia and New Zealand. Difference the incident schedule must preserve: Australia distinguishes a suspected-breach assessment period from notification once the statutory threshold is met. New Zealand requires notice as soon as practicable for a notifiable privacy breach; the regulator’s 72-hour expectation is guidance. Australian Privacy Act, Part IIIC; NZ Privacy Act, ss 112–115; OPC guidance.

Other applicable duties need their own entries, including Kenya’s computer-system and personal-data rules, designated critical-infrastructure duties, financial supervision and the relevant UAE or Saudi data regime. The customer cannot veto a report the provider is independently required to make. Nor should a dispute about fault delay factual cooperation, lawful containment or an initial notification that can later be supplemented.

Decide which work may start

For a customer-controlled target, accept only the operations supported by the customer’s rights, applicable platform permission and completed local conditions. For a third-party target, obtain the relevant grant or remove that target and its effects from the work. For a target without established authority, defer the affected activity while identifying the exact evidence needed. Refuse a demand to continue after withdrawal, conceal an incident or rely on an indemnity to justify known unauthorised access.

The final approval should identify accepted assets, responsible humans, scope version, data arrangements, incident contacts and open exclusions. It should also identify changes that require another legal review. Chile’s enacted privacy law sets general commencement on 1 December 2026. As at 18 November 2026, the parliamentary postponement text proposes 15 October 2027; rely on that later date for an engagement only after its promulgation and publication. Law 21.719, transitional provisions; Bill 18.623-07 legislative record.

Review the scope before testing starts

Bring the target inventory, authority evidence, platform conditions and proposed customer/provider terms to Licentium for a review of the permissions, local amendments and conditions required before activation.

Planning a launch somewhere new?

Prefer email? Write to hello@licentium.io. By sending, you accept our privacy policy.