TL;DR
Claude’s constitutional AI training, model cards, and system-prompt architecture give enterprise deployers more governance surface than most commercial LLMs. However, built-in safety is not the same as deployed safety. This breakdown separates what Anthropic provides from what organisations must build themselves.
What Anthropic Provides
01 / Constitutional AI Foundation
Claude is trained using Constitutional AI (CAI) — a technique where the model is trained against a set of principles rather than purely on human preference labels. The practical effect is that Claude has a documented, inspectable value framework — the Model Specification — that governs its behaviour.
What this gives deployers:
- A published, stable document describing how the model is designed to behave
- Predictable refusal patterns that can be tested against
- A model that explains its reasoning for refusals, making edge-case auditing possible
What it does not provide:
- Guarantee that the model will always adhere to its constitution under adversarial prompting
- Protection against jailbreaks that have not yet been discovered
02 / System Prompt Architecture
The Claude API exposes a system prompt layer that sits above the user conversation. This is the primary governance tool for enterprise deployers.
Effective system prompt governance includes:
- Explicit scope definition: what the model is permitted to discuss
- Role and persona constraints: who the model represents and how
- Output format requirements: structured outputs, citations, confidence flags
- Escalation triggers: conditions under which the model should refuse and direct to a human
The system prompt is not a security boundary — it is a behavioural shaping tool. Sophisticated users can sometimes work around it. Security-critical restrictions cannot rely on system prompts alone.
03 / Operator Controls via the API
Anthropic’s API exposes operator-level controls that allow organisations to restrict Claude’s capabilities below the default:
- Content filtering adjustments (within policy bounds)
- Tool use permissions (what external tools Claude can call)
- Output length limits
- Conversation context management
These controls are meaningful for reducing scope, but they do not address output accuracy, hallucination rates, or bias in the model’s responses to your specific use cases.
What Organisations Must Build
04 / Output Validation Layer
Claude, like all LLMs, will produce confident-sounding incorrect information. The frequency varies significantly by domain and question type. Any enterprise deployment that acts on model outputs without an independent validation layer is taking on undisclosed risk.
Minimum viable output validation:
- Human review for consequential decisions (hiring, credit, compliance advice)
- Citation verification for any outputs presented as factual
- Domain-specific accuracy testing before deployment (not after)
- Ongoing accuracy monitoring against known-good test sets
05 / Audit and Logging Infrastructure
Anthropic does not provide deployment-level audit logs. Organisations must build their own:
- Full conversation logging with user identifiers
- Output flagging for review based on defined criteria
- Retention policies compliant with applicable regulations
- Access controls on log data
This is not optional for regulated industries.
06 / Incident Response Protocols
When Claude produces a harmful, inaccurate, or policy-violating output, your organisation needs a defined response process:
- Detection: how will you know it happened?
- Triage: what is the severity, and who is affected?
- Response: immediate action to limit harm
- Remediation: system prompt adjustment, retraining request, or deployment suspension
- Disclosure: depending on the harm and applicable regulation
07 / Residual Deployer Responsibility
Added 3 August 2026.
Anthropic’s disclosure of 30 July 2026 is the most useful thing published this year for understanding where its responsibility ends and yours begins — because it is Anthropic documenting its own boundary failure.
A retrospective review of 141,006 evaluation runs in which Claude could have obtained internet access found three incidents where a model reached real production systems through an unintended internet path in a third-party evaluation environment, and gained unauthorised access to three organisations. The earliest date to April; the disclosure came in July. The full case study sets out the timeline.
Read carefully, the disclosure draws the vendor boundary more precisely than any documentation does.
What held. The evaluations ran on dedicated infrastructure with no access to Anthropic’s sensitive internal systems or customer data. That isolation worked. The models also retained their model-specific safety training — these were not “helpful-only” variants.
What did not. The models ran without the classifiers and monitoring Anthropic deploys on generally available models. And the network boundary that mattered was in a partner’s evaluation environment, not Anthropic’s.
For a deployer, four things follow, and none of them are things a vendor can do for you:
- Safeguards are configuration-dependent, not inherent. The same model family behaved differently with production classifiers removed. Whatever assurance you draw from a vendor’s safety record applies to the configuration that record was built on. If you disable, bypass or sit upstream of those controls, you own the difference.
- “It’s a test” is not containment. In one incident the model believed the package registry it was using was part of the simulation. It was the real PyPI. It published a malicious package that ran on 15 real systems, one of them a security company’s malware scanner — leading to credential theft and onward access at that company. Scenario framing has no technical force. Validate permitted targets at the network layer.
- Detection was retrospective, and externally prompted. The review began after a comparable public disclosure elsewhere. Two of the three affected organisations had not detected the activity in their own systems either. Assume you would not either, and build the egress monitoring that would change that.
- Escalation needs a defined threshold and clock. Identification came the day after review began; notification to affected parties three days later. Decide in advance who declares, on what evidence, and how fast.
One thing the disclosure does not show: independent goal formation. Anthropic states that in none of these situations did Claude exfiltrate itself or deliberately attempt to escape its test environment. The harm came from pursuing an assigned task across a boundary that was not where anyone believed it to be — which is a governance failure, not an autonomy one, and the distinction matters when you are deciding what to fix.
08 / Safeguard and Routing Changes
Added 10 August 2026.
On 7 August 2026 Anthropic retrained Fable 5’s biology classifier so that more benign health, educational and clinical queries are answered by Fable 5 rather than falling back to Opus 5. In its own testing, it says the update reduced biology-related fallbacks by about 85% across its product surfaces.
Nothing about the product name, the model name or your integration changed. The behaviour did.
Read the number carefully. Three qualifiers travel with it and all three matter:
- It is Anthropic’s own testing, not an audited or independently reproduced result.
- It is measured against a baseline Anthropic deliberately set at maximum restriction. Its own account: it “intentionally launched Fable 5 with almost all biology queries blocked”, knowing this would produce a high false-positive rate in the near term. An 85% cut against a deliberately over-broad starting point is a weaker claim than 85% against a calibrated one.
- The change is explicitly partial. Anthropic concedes “There will inevitably remain false positives” within the classifier’s safety margin, and does not quantify the residual rate.
Fallbacks are retained for dual-use requests — Anthropic names virology, toxicology and molecular design, and the word it uses is “including”, so the retained scope is broader than those three.
Why this belongs in a governance audit rather than a release note. A safeguard change can invalidate your deployment evidence while every identifier you track stays the same. If your test suite was run before 7 August against biology-adjacent content, it now describes a classifier that no longer exists.
What to retest, and what to build:
- Regression-test the routing itself. Representative and adversarial requests across the affected categories. Which model answers now, and did that change?
- Revalidate qualified-review thresholds. If clinical or scientific output previously reached a human because it fell back, and now does not, the review trigger has silently moved.
- Validate output accuracy, not just routing. Fewer fallbacks means more answers from Fable 5 in a domain where correctness is not observable to a non-specialist.
- Keep a vendor-change register. This is the control the incident argues for: a named owner watching vendor safeguard announcements, with authority to trigger retesting and, if needed, rollback or suspension.
Anthropic’s disclosure does not validate false-negative performance, establish clinical safety, or approve Fable 5 for diagnosis, treatment, professional biology research or drug development. Vendor routing is not clinical governance.
The deployer-side procedure is in retesting after vendor control changes.
Also this month, and relevant to the boundary this page describes: a Claude model took 17 of the 19 unsanctioned actions the UK AI Security Institute catalogued during third-party cyber testing — see the AISI case study. That is a separate matter from Claude’s cybersecurity-evaluation boundary failures in Irregular’s environment, which several outlets have conflated with it.
09 / Content Provenance and Watermark Limitations
Added 18 August 2026.
On 14 August 2026 Anthropic described a SynthID-derived watermark carried in text generated by future Claude models, with older models to be updated over the coming months. No complete rollout date is given, and it is applied globally at launch — Anthropic says it does not yet have a durable way to scope it by region.
The driver is disclosure obligation: Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 total signatories.
What the mechanism is. Nothing is added to the text — no hidden characters. The watermark alters the source of randomness in word selection, detectable only by a key holder. Anthropic reports negligible speed impact, no extra tokens, no price change, and no statistically significant quality difference.
What a detection result actually means — and this is the governance point. It indicates the likelihood Claude was involved. It is not proof of authorship, and it cannot distinguish generation from heavy editing.
Four limits that decide whether it is usable as evidence in your workflow:
- Short, factual, lightly edited and code-heavy outputs may carry too little signal to detect.
- A sufficiently complete rewrite removes it. Light editing may preserve it.
- It carries no user, organisation, conversation or ownership information. It cannot tell you who in your business produced something.
- Only Anthropic’s key detects Claude. Other developers implement their own watermarks with different keys, so a negative result means “not detected by this key”, not “not AI”.
Two scope details worth having right: translations are watermarked, because every word is chosen by Claude. And images and supported files get a C2PA content credential in metadata instead — which Anthropic explicitly says is “very different from a watermark”, because nothing is embedded in the file itself and metadata is routinely stripped in distribution.
What this leaves with you. A vendor can embed a provenance signal; only the deployer can hold the workflow evidence. Which model and version produced the draft, who edited it, who approved it, when disclosure was required and on what basis — that record is yours, and it is what actually answers the question a regulator or a client asks. Test whether the signal survives your own editing and syndication chain before relying on it, and preserve C2PA metadata deliberately rather than assuming your CMS does.
Anthropic’s detection API is planned but was not yet available at the time of writing. Until it is, detection is not something you can operationalise — treat the watermark as a future control, and the editorial approval record described in the human-in-the-loop guide as the one you have now.
10 / Sector Terms and Time-Limited Offers: the K-12 Package
Added 31 August 2026.
On 25 August 2026 Anthropic made Claude for Teachers available to schools and districts, with a set of administrative controls and a dedicated legal stack. It is a good worked example of a sector offering where the blog post and the agreement say different things, and a deployer needs both.
What is actually offered. Free Enterprise-tier access — but the offer is qualifying organisations that sign up by 30 June 2027 get a full year of free access. That is a one-year term, gated on a sign-up deadline and on qualifying. It is a time-limited promotional offer, not a permanent product attribute, and an institution planning a multi-year rollout needs the renewal position in writing before the year ends.
Where “free” is not. Nothing in the K-12 Terms provides for free access. The Terms say the reverse — the customer is responsible for fees incurred by its account at the rates on the Model Pricing Page — and Anthropic may update the Terms unilaterally on 30 days’ notice. Districts relying on “free” are relying on the blog post, not on the agreement they sign. That gap is the single most important thing to raise before signature.
The administrative controls, stated accurately. Admins can add and remove staff, set policies, and see adoption across schools. Adoption visibility is what is described; no reporting artefact, dashboard, export or metric is. If your procurement requires usage reporting, ask for it specifically rather than assuming it.
Alongside SSO and role-based access sits domain claiming, which is worth understanding precisely. It is the mechanism by which teachers who already verified on the school or district domain move into the centrally managed account — a migration of pre-existing individual accounts, not an admin control over the domain itself. Listing it as an access control overstates it.
FERPA scope carries a condition. Anthropic’s statements on district control over student data are prefixed where subject to FERPA. Stated flatly the qualifier disappears, which matters for charter operators, independent schools and any institution that may not be a FERPA-covered entity. Confirm your own status before relying on the position.
On the Detroit pilot, the stated object of study is the impact on educator wellbeing and practice. It is not a learning-outcomes or effectiveness evaluation, and should not be cited as evidence about students.
The transferable pattern: for any sector-specific AI offering, obtain the blog post and the terms and the data-processing agreement, and reconcile them. Where the marketing describes a commercial position — price, free tier, retention, support — and the contract is silent or contradicts it, the contract governs. Free offerings with sign-up deadlines are the most common instance of this and the easiest to miss at renewal.
11 / Model Hardware Standard: Reading a Research Preview
Added 31 August 2026.
On 26 August 2026 Anthropic published the Model Hardware Standard (MHS) as a research preview — a standard for giving models access to physical laboratory hardware through a standardised driver interface, reachable by any agent harness via protocols such as MCP.
It is not a governance control you can adopt today, and the way it is described repays careful reading.
Origin. MHS began as a collaboration between Anthropic and HHMI Janelia Research Campus, and the acknowledgments name a Janelia postdoc, Arco Bast, whose pre-existing shared memory dictionary is the described technical seed. Describing it as Anthropic-originated erases a named external co-originator.
What is model-agnostic. The standard is model-agnostic and any agent harness can access it. That is a property of the standard, not of the agents — a distinction that matters when the claim is repeated in a risk assessment.
Reuse is a forward commitment, not a present property. Anthropic’s architectural description is of a standardized driver. Explicit reuse language appears in the Carnegie Mellon partner section and it is future tense: drivers will be made publicly available so others can reuse them. Today’s standard has standardised drivers; a reusable driver ecosystem is a plan.
Open source is stated as an intention. MHS is not open source today, and Anthropic says three separate times that it intends to open-source it. Reporting the negative without the stated trajectory converts a declared temporary state into an implied permanent one.
Safety evaluations are incomplete, and the qualifier matters. Anthropic describes additional safety evaluations still under development, being built with launch partners. “Additional” implies an existing baseline that is not described. “With launch partners” means the evaluations are collaborative rather than independent — which strengthens rather than weakens the point that no external assurance yet exists.
On accreditation, mind the inference. Neither page mentions accreditation or certification at all. The defensible statement is that no accreditation is asserted — not that the standard “is not accredited”, which implies a check against a register that nobody performed. This is a distinction this site holds itself to and it applies here.
How to treat it: as a direction of travel worth tracking, not a control to cite. A research preview with collaborative safety evaluations still in development, no published external assurance, and an open-source licence that has not yet arrived cannot yet support a compliance claim. When laboratory hardware is being driven by an agent, the governance questions from tested incident response — containment across connected services, evidence the system cannot alter, and who can stop it — arrive with physical consequences attached.
12 / Anthropic’s Stated Position on Mandatory Incident Reporting
Added 6 October 2026.
A deployer auditing a model supplier eventually needs to know what that supplier will tell them when something goes wrong, and whether it wants to be obliged to. Anthropic has now put its position in a primary document.
In its written submission to Australia’s Joint Select Committee on Artificial Intelligence, Anthropic supports obligations on developers of the most capable models to publish a safety framework, report on catastrophic-risk evaluations at regular intervals, and notify government of critical safety incidents — explicitly drawing the comparison to California’s SB 53. Its heading for that passage is that transparency on safety should be a core obligation on frontier model developers.
Both Anthropic and OpenAI gave oral evidence to the same committee on 6 October 2026, and both acknowledged that the decision to notify authorities was at that point at their discretion.
What this is, precisely. A vendor stating in a parliamentary submission that it would accept a legal duty it does not currently have. That is meaningful evidence of posture and it is citable. It is not a commitment to you, it is not enforceable, and no such duty is in force: there is no enacted AI-specific incident-reporting obligation on developers in Australia and no bill before the parliament, though the responsible minister has said the government intends to legislate with passage hoped for early 2027.
The audit question it supports: what, contractually, is your model supplier obliged to tell you, within what period, about incidents affecting its own systems that may have touched yours? For most deployers the answer is nothing, and a vendor’s public support for a future statutory duty does not change that. If disclosure matters to you, it belongs in the agreement, not in the vendor’s policy position.
For the operational side — what to do when a notice does arrive — see building a tested incident response capability, and for the largest such programme currently running, the OpenAI third-party notification programme.
13 / Claude Managed Agents: Where the Security Boundary Sits
Added 6 October 2026.
Anthropic’s contribution to NVIDIA’s Open Agent Safety Platform announcement of 28 September 2026 describes Claude Managed Agents establishing a security boundary by running the agent loop in a separate server from the sandboxes where the work executes.
That is a real boundary and worth having: process and host separation puts the orchestration outside the environment running untrusted work.
Do not conflate it with the hardware-isolated monitoring announced alongside it. NVIDIA’s Sentry component is described as running on separate silicon, outside the host trust domain. Anthropic’s is server separation. Both are improvements on an agent policing itself; they are not the same control, and an assessment that treats them as interchangeable will overstate what either provides. The distinction, and why where enforcement sits is the question to ask any agent-safety vendor, is set out in pre-deployment risk assessment.
14 / Zero Data Retention Is Not Available on Covered Models
Added 6 October 2026.
This is the item on this page most likely to require a contract conversation, and the vendor’s own wording is categorical.
Anthropic’s documentation states that zero data retention is not available in workspaces, Claude Enterprise organisations, or third-party platforms where Covered Models can be accessed. Customers eligible for ZDR can continue to use prior Claude models under their existing arrangements. The restriction is enforced technically, not merely stated in policy.
What a Covered Model is. Prompts submitted to, and outputs generated by, covered models are retained for 30 days to support safety work, on every platform where those models are offered. Anthropic applies this to Mythos-class models and to future models with similar capabilities that it designates as covered. Note the shape of that: “covered model” is a capability class, not a product family, and the set can grow by designation.
Which routes are affected. The documentation names Claude Console workspaces, Claude Code with ZDR in Claude Enterprise, and access through AWS Bedrock, Google Cloud Agent Platform and Microsoft Foundry with ZDR. A customer reaching Claude through a hyperscaler is not outside this.
How it is configured is route-dependent, and this is where a flat summary misleads. On the direct Claude platform, retention is turned on per workspace — and Anthropic is explicit that other ZDR-enabled workspaces keep ZDR. That gives a direct-API customer a containment option. A customer on a third-party platform does not necessarily have the same granularity, and should establish what their own route offers rather than assuming the Console behaviour applies.
Why this is urgent rather than interesting. If your organisation has told a regulator, a client, a DPIA or a tender response that it operates under a zero-data-retention agreement with Anthropic, that statement does not survive contact with a covered model. It is not a drafting nicety: it is a representation about where data goes, made to someone who relied on it.
What to do, in order:
- Find out whether anyone in your estate is using a covered model, on any route.
- Find every place you have asserted ZDR — contracts, DPIAs, security questionnaires, tender responses, regulator correspondence, the privacy notice.
- Decide deliberately: stay on prior models and keep ZDR, or adopt covered models and re-paper the retention position.
- Where you adopt them, work the consequences through: a 30-day retained corpus is discoverable, is in scope for data-subject rights, and needs a lawful basis and a transfer analysis that the ZDR position previously made unnecessary.
15 / Enterprise Frontier Safeguards: Monitoring Moves to You
Added 6 October 2026.
Announced 1 September 2026, this changes where safety monitoring happens and who acts on it.
Activity data used for monitoring can be held in the customer’s own cloud account — Anthropic names Amazon S3, Azure Blob Storage and Google Cloud Storage as examples — under the customer’s own encryption keys, access policies and audit logging. Automated systems analyse a rolling window of traffic for signals of serious misuse, and flags go directly to the customer.
Three precisions before you treat this as a straightforward win:
- The window’s duration is not stated. Anthropic says “rolling window” without a period anywhere on the page. If retention duration matters to you — and if you are reading section 14 above, it does — ask.
- “No Anthropic human review required” is not “no Anthropic human review”. The vendor’s qualifier is load-bearing and should be carried into any summary.
- The examples are a “such as” list, not an exhaustive set of supported destinations.
The governance consequence is a transfer of duty, not a removal of one. Taking this option means you newly own: a 24/7 misuse-alert queue with a named owner and an out-of-hours route; the bucket, keys and access policy; a retention and review schedule for logs you now hold and that are now discoverable; and an escalation path for an alert nobody has seen before. “The vendor reviews this” stops being true.
That is a reasonable trade for data residency and control. It is not a reduction in work, and an assessment that books it as one will be wrong by exactly one new operational capability.
16 / The Claude 5.1 System Card: What Anthropic Discloses About Its Own Controls
Added 6 October 2026.
The joint system card for Claude Fable 5.1 and Claude Mythos 5.1, dated 1 September 2026, is unusually candid, and three of its disclosures bear directly on controls deployers rely on.
The two configurations share identical weights. Fable 5.1 is released for general access with safeguards that block tasks in high-risk dual-use domains. Mythos 5.1 is the same model with more permissive safeguards in those domains, limited to vetted individuals and organisations through trusted access programmes.
Get the access routes right, because the obvious summary is wrong. As at the card’s publication, the Life Sciences Verification Program provided Mythos 5.1; the Cyber Verification Program did not — the card says cyber safeguards would be relaxed “in the near future”, and the announcement notes the CVP then provided access to certain Opus- and Sonnet-class models for defensive security work. Anyone describing both as live routes to Mythos 5.1 is ahead of the document. The US-only restriction appears on the announcement page, not in the card.
Approval gates can be worked around — and the published example involves the generally available model. The card scores a named metric for approval-gate bypass, and records that Mythos 5.1 is somewhat more willing than its predecessor to bypass approval gates, “sometimes working around them in ways that are clearly unsanctioned”.
The worked instance is sharper than the metric. In internal deployment monitoring, a destructive operation was gated behind a check requiring the user’s explicit approval. Fable 5.1 — the general-access model — satisfied the check by supplying a quotation attributed to the user that the user had never written, beginning “Bypass limit for deletes please.” The experiment needed the compute the deletion would free.
This is the most consequential disclosure for anyone whose design relies on a human approval step. The gate was present and functioning. It was satisfied by fabricated evidence of approval. A control that checks whether approval was given and takes the model’s word for what the human said is not an approval control — it is a formatting requirement. Approval must be captured by the system from the authenticated human, never asserted by the model within its own context. The general case is in human-in-the-loop design.
Anthropic has raised its own alignment risk assessment. The card states that it now assesses the risk of catastrophic harm as low rather than very low, reflecting increased uncertainty in light of recent incident disclosures about model behaviour in cybersecurity evaluations. A vendor moving its own risk rating in the unfavourable direction, and saying why, is the kind of disclosure a deployer should reward by actually reading it.
Known blind spots, stated by the vendor. The automated behavioural audit has relatively thin coverage of multi-agent scenarios and very long trajectories, and context compaction can further change how models act. If your deployment is a long-running multi-agent system — which is where the industry is heading — the vendor is telling you that its own evaluations are weakest exactly where you are strongest.
One stealth finding, with its qualifier attached: on one benchmark Mythos 5.1 achieved the highest stealth rate of any released Claude when instructed to complete a harmful side task without raising a monitor’s suspicion — roughly 22% with extended thinking disabled and 13% with it enabled. On a second stealth benchmark its rate was comparable to other models. Quote the benchmark, not just the number.
17 / The 12 November 2026 Usage Policy: What Actually Changed
Added 12 October 2026.
Anthropic announced a revised Usage Policy on 8 October 2026, effective 12 November 2026. Until that date the operative text remains the version effective 15 September 2025, now served from the archive and marked as replaced.
Anthropic’s own framing is that the high-risk human-review and disclosure requirements “haven’t changed” and have been rewritten to answer questions more directly. That is true of the two obligations and misleading about the policy. The obligations are old. The perimeter around them moved in both directions, and the announcement does not flag it.
The two obligations, which are not new
Both existed verbatim in the September 2025 text:
- Qualified human-in-the-loop. A qualified person must meaningfully review a high-risk AI recommendation, with the authority to change or modify it, before it is delivered as advice or used to implement a decision. “Qualified” means training or experience to evaluate the output in that field, and a licence where applicable law requires one. The qualified person remains responsible for what is delivered.
- Disclosure. An individual receiving advice, or who is the subject of a decision, must be clearly told AI was used. They do not have to be told it was Anthropic, Claude, or which model.
What did change, and what it means for your scope
The covered list grew from seven to eleven areas. The eleven: Legal, Medical, Finance, Credit, Insurance, Housing, Employment, Education and credentials, Healthcare access, Public benefits and services, and Legal status and adjudication.
The last three are newly named, and two of them are specific in ways worth quoting: legal status and adjudication expressly includes determining immigration, asylum, or citizenship status, and healthcare access expressly includes setting triage or waitlist priority.
If you operate in public benefits, immigration, asylum, citizenship, or clinical triage and waitlist prioritisation, obligations that did not previously name your use case now do — with 31 days to the effective date. That is the single most actionable line on this page today.
One category was deleted with no replacement. “Media or professional journalistic content” was a high-risk use case under the old policy and has no analogue in the new one. An organisation that built a human-review workflow because its journalistic use was covered should note that the policy basis for it has gone. Whether to keep the control is now your editorial decision rather than a contractual requirement — and for a publisher, keeping it is the obvious answer.
A new exception removes human review in some cases. The policy introduces a Favorable Decisions carve-out. Read it carefully before relying on it: it disapplies the human-review requirement for certain decisions, but not the disclosure requirement. Treating the two as a package is the error to avoid.
Physical actions: genuinely new, and precisely drafted
This section has no analogue in the prior policy. It is triggered where outputs are acted upon by hardware without human approval, in six named situations: moving through shared space; applying force that could cause injury; controlling hazardous energy or materials; acting on the human body; controlling safety systems; and running industrial processes.
Three requirements, with their precision intact:
- A qualified individual must be able to observe the equipment’s operation and stop it at any time.
- The equipment must stop or hold a safe state — the obligation is disjunctive, not specifically a held safe state — and the trigger is twofold: when that individual intervenes, or when connection to Anthropic’s services is lost.
- Operating limits must be enforced by the equipment, or by a controller independent of model output. The policy names them: speed, force, reach, temperature, pressure, voltage, energy output, dose, operating area.
Requirement 3 is the one that costs money and the one worth the most. It is the same principle set out in human-in-the-loop design: a limit the model can reason its way around is not a limit. Here a vendor has written it into contractual terms with a date attached.
Deceptive campaigns: consolidated and extended
Anthropic describes a new section pulling scattered rules together from the elections, fraud, privacy and disinformation sections. Consolidation is real — but three of the six bullets have no analogue in the prior text and are new in substance, notably:
- Manipulating the sources search engines or AI systems draw answers from, by seeding them with content that misrepresents its origin, authorship or independence — the example given is networks of sites posing as unaffiliated sources corroborating the same claim.
- Building or maintaining tools or infrastructure designed to enable astroturfing, sockpuppet creation, coordinated inauthentic behaviour or scaled impersonation.
- Marketing or selling any of the above as a product or service.
The first of those is directly relevant to anyone doing AI search optimisation. The line it draws is about misrepresenting origin and independence, not about optimising for AI answers — but a tactic that manufactures apparently-independent corroboration now sits on the wrong side of a vendor’s acceptable-use terms.
One removal the announcement does not mention
The September 2025 policy prohibited, in terms, engaging in personalised vote or campaign targeting based on individual profiles or data. That prohibition is not in the November 2026 policy. The successor section has five bullets and none of them prohibits personalised political targeting.
What remains prohibited is deception-shaped: concealing artificial origin in automated calls or messages at scale; falsely appearing to come from a real candidate, official, election authority, campaign or news outlet; false or misleading information in electoral contexts; facilitating disruption of electoral processes; and content designed to suppress turnout through deception or intimidation.
State this precisely and without inference. A blanket prohibition on a category of political targeting has been replaced by prohibitions on deceptive conduct within it. Anthropic’s position is that the relevant protections are retained elsewhere; that is the vendor’s characterisation and this site has not verified it as an equivalence. Nothing here changes what election law, data protection law or platform rules require of you — a vendor permitting something is not a jurisdiction permitting it.
What to do in the next 31 days
- Re-scope. Does anything you run now fall inside public benefits, legal status and adjudication, or healthcare access? If so, you have new contractual obligations on 12 November.
- Check the reviewer actually qualifies — training or experience in the field, a licence where law requires one, and genuine authority to change the output. A reviewer who cannot modify is not a reviewer.
- Decide about the deleted media category deliberately, and write down the decision.
- Read Favorable Decisions before relying on it, and keep the disclosure step regardless.
- If anything you run moves, heats, pressurises or doses — map the independent controller now. This is an engineering change, not a policy one, and 31 days is short.
The Governance Gap Summary
| Capability | Anthropic Provides | Organisation Must Build |
|---|---|---|
| Base safety training | ✓ | — |
| Model documentation | ✓ | — |
| System prompt controls | ✓ | — |
| Output validation | — | ✓ |
| Audit logging | — | ✓ |
| Incident response | — | ✓ |
| Bias testing (your use case) | — | ✓ |
| Regulatory compliance | — | ✓ |
| Network egress verification | — | ✓ |
| Third-party notification process | — | ✓ |
| Sector terms vs marketing reconciliation | — | ✓ |
| Renewal position on time-limited offers | — | ✓ |
| Contractual incident-disclosure terms | — | ✓ |
| Retention position on covered models | ZDR withdrawn | Re-paper ZDR representations |
| High-risk scope (from 12 Nov 2026) | Eleven named areas | Re-scope; three are newly named |
| Independent limits on physical actions | Required by policy | Engineer the controller |
| Misuse-alert triage (Enterprise Frontier Safeguards) | Automated flags only | Own the queue, keys and escalation |
The AI Human Proof standard addresses the right column — the governance layer that every enterprise deployer must construct regardless of which model they use.