Examples
Examples from published documents
- Anthropic's misalignment assessment covers eight concrete pathways, among them "Code backdoors to help future misaligned models", "Self-exfiltration and autonomous operation" and "Persistent rogue internal deployment" (Anthropic Risk Report, August 2026, section 2.2.1)
- OpenAI's safeguards for Astra address two risk pathways: "(1) a malicious actor using Astra to develop novel exploits or carry out end to end attacks against hardened critical systems or (2) the model itself causing cyber harm when taking an unauthorized (or misaligned) action" (OpenAI, Path to Astra, section 10.2)
- Meta requires that "It must be possible to identify a causal pathway for the catastrophic outcome, and to define one or more simulatable threat scenarios along that pathway" before an outcome enters its framework (Meta Advanced AI Scaling Framework v2, section 3.2)
Counter-examples
Not to be confused with
- A risk domain such as CBRN or cyber offence is not a risk pathway: it groups many pathways under one heading. xAI's framework calls its risk domains "principal pathways", a broader use of the same word.
- A threat scenario is not the pathway itself: the scenario is one worked case (actor, target, steps), while the pathway is the causal route that such cases follow.
Related terms
Related AI safety terms
Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.
- Self-exfiltration
- Self-exfiltration is an AI model copying or moving its own weights out of its developer's secure storage. The EU Code of Practice covers transfer of weights "by the model itself and/or an unauthorised actor"; Anthropic lists self-exfiltration as a priority risk pathway.
- Source: EU GPAI Code of Practice, Safety and Security Chapter
- Subterfuge
- Subterfuge is the word Anthropic's Responsible Scaling Policy v3.4 uses for covert, deceptive conduct by an AI system. Its misalignment threshold covers systems with "moderate capacity for autonomous, goal-directed operation and subterfuge" that could carry out sabotage.
- Source: Anthropic Responsible Scaling Policy v3.4
- Scheming
- Scheming, which Google DeepMind's Gemini 3.7 Flash report says is "also known as deceptive alignment", is a model knowingly and covertly pursuing objectives misaligned with its developer's intention. OpenAI's Preparedness Framework lists deception and scheming as misaligned behaviours.
- Source: Gemini 3.7 Flash FSF report
- Stealth and situational awareness
- Stealth and situational awareness are the two capabilities in Google DeepMind's misalignment threshold: situational awareness is the "ability to discover and use relevant details of its deployment setting", stealth the "ability to circumvent basic oversight mechanisms".
- Source: Google DeepMind Frontier Safety Framework v3.1
- Covert capabilities
- Covert capabilities, in Anthropic's August 2026 Risk Report, are "the ability for a model to intentionally undermine the various oversight mechanisms we rely on to evaluate and mitigate model behavior", split into opaque reasoning, secret keeping, action obfuscation and selective action.
- Source: Anthropic Risk Report, August 2026
Source of record
Where this definition comes from
Anthropic Risk Report, August 2026, §2.2.1, §2.13
“"Our risk assessment focuses on the following concrete pathways": "1. Broad/diffuse sandbagging on safety-relevant R&D, 2. Targeted undermining of safety-relevant R&D, 3. Code backdoors to help future misaligned models, 4. Poisoning the training data of future models, 5. Self-exfiltration and autonomous operation, 6. Persistent rogue internal deployment, 7. Undermining R&D within other high-resource AI developers, and 8. Undermining decisions within major governments" (§2.2.1). "We aren't able to defend the choice of these pathways rigorously" (§2.13).”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdfGoogle DeepMind Frontier Safety Framework v3.1, s.1.2
“"CCLs are determined by identifying and analyzing the main foreseeable paths through which a model could result in severe harm: we then define the CCLs as the minimal set of capabilities a model must possess to do so." (s.1.2)”
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3-1.pdf
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Anthropic Risk Report, August 2026 | “"Our risk assessment focuses on the following concrete pathways": "1. Broad/diffuse sandbagging on safety-relevant R&D, 2. Targeted undermining of safety-relevant R&D, 3. Code backdoors to help future misaligned models, 4. Poisoning the training data of future models, 5. Self-exfiltration and autonomous operation, 6. Persistent rogue internal deployment, 7. Undermining R&D within other high-resource AI developers, and 8. Undermining decisions within major governments" (§2.2.1). "We aren't able to defend the choice of these pathways rigorously" (§2.13).” | exactEQ confidence: high | Anthropic Risk Report, August 2026 |
| OpenAIShadow mapping Preparedness Framework v2 / Path to Astra | “Safeguards Report contents: "Identified ways a risk of severe harm can be realized for the given deployment, each mapped to the associated security controls and safeguards" (PF §4.2). Astra: "Risk pathways (cyber)": "(1) a malicious actor using Astra to develop novel exploits ... or (2) the model itself causing cyber harm when taking an unauthorized (or misaligned) action." (s.10.2)” | closeCL confidence: high | OpenAI Preparedness Framework v2 |
| Google DeepMindShadow mapping Frontier Safety Framework v3.1 / Gemini 3.7 Flash FSF report | “FSF v3.1: "CCLs are determined by identifying and analyzing the main foreseeable paths through which a model could result in severe harm" (s.1.2). The Gemini 3.7 Flash FSF report restates this as "could cause severe harm" (report p.2).” | closeCL confidence: high | Google DeepMind Frontier Safety Framework v3.1 |
| xAIShadow mapping xAI Frontier AI Framework (30 Jun 2026) | “"These risk domains describe principal pathways through which severe or systemic harm may arise." (s.2.1). Here "pathway" names the risk domain itself, not a finer-grained causal route beneath it.” PDF metadata /Title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC"; no xAI statement disambiguating draft vs. final was found. | noneFF confidence: high | xAI Frontier AI Framework (30 Jun 2026, draft-marked) |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “Threat modeling "identifies the potential causal pathways for realizing the catastrophic outcome" (Appendix I).” | closeCL confidence: high | Meta Advanced AI Scaling Framework v2 |
| California SB 53Shadow mapping California SB 53 | “Catastrophic risk arises from "a single incident involving a frontier model doing any of the following: (A) ... (B) ... (C) ..." (22757.11(c)) — a broader, outcome-defined construct than a specific causal pathway.” | broadBR confidence: medium | California SB 53 |
| Frontier Model ForumShadow mapping FMF Risk Taxonomy and Thresholds | “Threat modeling includes "mapping the potential pathways to those outcomes"; "Credibility: There is a credible pathway to extreme harm." (s1.3, s2.1)” | closeCL confidence: high | FMF Risk Taxonomy and Thresholds |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “"systemic risk modelling": "a structured process aimed at specifying pathways through which a systemic risk stemming from a model might materialise; often used interchangeably with the term 'threat modelling'" (Glossary; see Measure 3.3).” | closeCL confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
What do these codes mean?
- exact
- The source term is equivalent to this element
- close
- The source term is close but not equivalent to this element
- broad
- The source term is broader than this element
- narrow
- The source term is narrower than this element
- none
- No mapping claim — used for false-friend and declared-but-undefined rows
- EQ
- Equivalent
- CL
- Close
- BR
- Source is broader than the element
- NR
- Source is narrower than the element
- FF
- False friend — same or similar label, different meaning
- DU
- Declared but undefined by the source
- UV
- Unverified
Related, not mapped
Pointers that are not crosswalk claims
These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.
- Anthropic
Pathway 6, "Persistent rogue internal deployment", is defined and assessed in the Risk Report (§2.11.3). The term rogue internal deployment is defined under the NIKOLAI element internal-deployment-risk (track N6); this element records it only as one pathway in Anthropic's list.
Anthropic Risk Report, August 2026
Divergence
Where sources materially disagree
xAI's Frontier AI Framework uses "pathway" as a synonym for its risk domains themselves (s.2.1), not as a finer-grained causal route beneath a domain/threat-model as in Anthropic's eight priority risk pathways, OpenAI's Astra risk pathways, or Meta's causal pathways — a false-friend label collision. SB 53 uses the same word for its three incident mechanisms.







