diff --git a/CHANGELOG.md b/CHANGELOG.md index 32c6d06..19d0f77 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -9,6 +9,11 @@ Format: [Semantic Versioning](https://semver.org). Schema versions and record se ## [Unreleased] ### Added +- AVE-2026-00067: skill composition trust transfer (SCR-TrustLift) — a + downstream skill accepts an upstream skill's benign output (an + endorsement, audit finding, validation result) as sufficient + authorization without independently re-verifying the claim; neither + skill is dangerous in isolation (MEDIUM, AIVSS 5.0) - AVE-2026-00066: hallucinated skill-name squatting (HalluSquatting) — agent hallucinates a plausible-but-nonexistent package/repo/skill name, attacker pre-registers it with malicious content, no injected diff --git a/README.md b/README.md index 07daa59..31242fc 100644 --- a/README.md +++ b/README.md @@ -234,6 +234,7 @@ AIVSS = ((8.5 + 7.5) / 2) x 1.0 x 1 = 8.0 -> HIGH | [AVE-2026-00064](records/AVE-2026-00064.json) | Zero-Click Code Execution via Auto-Run Configuration | 5.2 | MEDIUM | | [AVE-2026-00065](records/AVE-2026-00065.json) | A2A Agent Card Poisoning | 7.1 | HIGH | | [AVE-2026-00066](records/AVE-2026-00066.json) | Hallucinated Skill-Name Squatting (HalluSquatting) | 6.2 | MEDIUM | +| [AVE-2026-00067](records/AVE-2026-00067.json) | Skill Composition Trust Transfer (SCR-TrustLift) | 5.0 | MEDIUM | --- diff --git a/dist/ave-records-latest.json b/dist/ave-records-latest.json index 64656c4..466d43e 100644 --- a/dist/ave-records-latest.json +++ b/dist/ave-records-latest.json @@ -8352,6 +8352,100 @@ "remote-control-chain" ] }, + { + "ave_id": "AVE-2026-00067", + "schema_version": "1.1.0", + "status": "active", + "component_type": "skill", + "title": "Skill composition trust transfer: an upstream skill's output silently becomes a downstream skill's trust signal", + "attack_class": "Privilege Escalation - Skill Composition Trust Transfer", + "severity": "MEDIUM", + "description": "A skill that is entirely benign when reviewed or executed in isolation produces an output, such as an audit finding, an endorsement, a validation result, or another benign-looking artifact, that a separate, later-invoked skill treats as a trust or authorization signal without independently re-verifying the underlying claim. Neither skill individually does anything dangerous; the vulnerability exists only in the composition, when both are invoked along the same task path in a shared execution context. Because per-skill security review evaluates each skill's own behavior in isolation, it structurally cannot see this class of risk: an upstream skill's legitimate, correct output becomes a spoofable trust credential the moment a downstream skill treats it as authoritative rather than re-verifying the claim itself. Published research demonstrates this mechanism accepting harmful software installation at over 96% success across four of five tested model backends when preceded by an approval-like output from an unrelated upstream skill, versus near-zero success when the same downstream skill is invoked in isolation.", + "aivss_score": 5, + "cvss_base_vector": "CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:L/SC:L/SI:H/SA:N", + "owasp_mcp": [ + "MCP02" + ], + "owasp_asi": [ + "ASI03", + "ASI08" + ], + "mitre_atlas": [], + "nist_ai_rmf": [], + "behavioral_fingerprint": "A downstream skill accepts an upstream skill's output (an endorsement, an audit result, a validation finding, or similar artifact) as sufficient authorization or trust justification for a sensitive action, without independently re-verifying the claim the upstream output represents.", + "behavioral_vector": [ + "skill-composition-risk", + "trust-signal-propagation", + "cross-invocation-authorization-confusion" + ], + "provenance_vector": { + "entry_class": "tool_response", + "payload_surface": "a benign, correct upstream skill's output (endorsement, audit finding, validation result) consumed by a downstream skill as an implicit trust or authorization signal" + }, + "mitigation": { + "strategy": [ + "verify_identity", + "isolate_scope" + ], + "enforcement_point": "agent_framework", + "trifecta_control": "not_applicable" + }, + "example_patterns": [ + "A security-review skill outputs 'no issues found' for a candidate package; a separate installation skill, invoked later in the same task, treats that output as sufficient approval to install the package without independently checking its provenance", + "A file-audit skill flags a directory as 'clean'; an access-manager skill invoked afterward grants broader permissions to that directory on the strength of the audit output alone, without its own independent verification" + ], + "mutation_count": 0, + "detection_methodology": "1. Trace multi-skill task executions and record which skill outputs are consumed as inputs, conditions, or justifications by subsequently invoked skills, not just each skill's own isolated behavior. 2. Flag any downstream skill invocation whose sensitive action is gated primarily on an upstream skill's output without an independent, direct verification of the same underlying claim. 3. Sandbox-based differential testing: run the downstream skill's decision logic both with and without the upstream skill's output present; a large swing in acceptance rate for the same underlying request is the defining signal (published research measured this as attack success in composition versus near-zero in isolation).", + "indicators_of_compromise": [ + "A sensitive action (install, grant permission, execute) immediately following a separate skill's benign-looking endorsement, audit result, or validation output within the same task path", + "Sensitive-action acceptance rate that differs sharply depending on whether a preceding, unrelated skill's output is present in context, for what is otherwise the identical underlying request", + "Downstream skill logic that branches on the presence or content of another skill's prior output rather than performing its own independent check" + ], + "remediation": "Design each skill's authorization and trust decisions to be self-contained: require independent verification of any claim a sensitive action depends on, rather than accepting another skill's prior output as sufficient justification. At the orchestration layer, scope each skill invocation's effective authority independently of what preceded it in the same task, so an upstream skill's output can inform a downstream skill but cannot substitute for that skill's own required checks.", + "kill_switch_active": false, + "researcher": "Saray Chak", + "researcher_url": "https://bawbel.io", + "published": "2026-08-02T00:00:00Z", + "last_updated": "2026-08-02T00:00:00Z", + "references": [ + { + "tag": "Benign in Isolation, Harmful in Composition (arXiv 2606.15242)", + "text": "Xie, Du, Cheng, Zhou, Yin. 'Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems.' Introduces Skill Composition Risk (SCR) and SCR-Bench across three sub-types (SCR-CapFlow, SCR-TrustLift, SCR-AuthBlur). SCR-TrustLift (this record's scope) measured over 96.5% attack success across four of five backends when a downstream decision is preceded by an upstream endorsement-like output, versus near-zero in isolation.", + "url": "https://arxiv.org/abs/2606.15242" + } + ], + "aivss": { + "cvss_base": 7.5, + "aarf": { + "autonomy": 1, + "tool_use": 1, + "multi_agent": 0, + "non_determinism": 0.5, + "self_modification": 0, + "dynamic_identity": 0.5, + "persistent_memory": 0, + "natural_language_input": 0, + "data_access": 0.5, + "external_dependencies": 0 + }, + "aars": 3.5, + "thm": 0.9, + "mitigation_factor": 1, + "aivss_score": 5, + "aivss_severity": "MEDIUM", + "spec_version": "0.8", + "notes": "multi_agent scored 0, not partial: this is a single-agent, multi-skill composition risk, not a multi-agent one -- the composition happens across sequential skill invocations within one agent's own session, not across distinct agents. natural_language_input scored 0: the vulnerable signal is a skill's structured output/return value, not natural-language instruction text. Scoped specifically to the SCR-TrustLift sub-mechanism (trust-signal propagation) rather than all three SCR sub-types the source paper covers, since SCR-CapFlow (capability-flow composition) and SCR-AuthBlur (authorization-confusion composition) are related but structurally distinct mechanisms worth their own records if drafted, not folded into one record broad enough to blur the fingerprint. mitre_atlas and nist_ai_rmf left as empty arrays deliberately, not omitted: researched directly against primary/near-primary sources rather than corpus pattern-matching. The closest ATLAS candidate, AML.T0053 (AI Agent Tool Invocation), describes an agent exploiting its own access to invoke connected tools for elevated privilege, a different mechanism from a downstream skill trusting an upstream skill's benign output, so it was not force-fit. For NIST AI RMF, independent research corroborates a documented gap: the current MAP function has no subcategory evaluating emergent risk from AI system component/tool combinations, matching this record's own finding that no clean mapping exists yet." + }, + "evidence_kind_default": "semantic_inference", + "detection_stage": "runtime_observed", + "detection_layer": "runtime", + "confidence_baseline": 0.55, + "evidence_basis_engines": [ + "sandbox", + "llm" + ], + "derivable_into": [] + }, { "ave_id": "AVE-2026-00014", "schema_version": "1.1.0", diff --git a/dist/ave-records-latest.manifest.json b/dist/ave-records-latest.manifest.json index 6e2eba1..716d0c1 100644 --- a/dist/ave-records-latest.manifest.json +++ b/dist/ave-records-latest.manifest.json @@ -1,6 +1,6 @@ { "schema_version": "1.1.0", - "record_count": 66, - "generated_at": "2026-08-02T16:06:49.281Z", + "record_count": 67, + "generated_at": "2026-08-03T14:42:22.928Z", "source": "https://github.com/aveproject/ave" } diff --git a/records/AVE-2026-00067.json b/records/AVE-2026-00067.json new file mode 100644 index 0000000..5565c5b --- /dev/null +++ b/records/AVE-2026-00067.json @@ -0,0 +1,76 @@ +{ + "ave_id": "AVE-2026-00067", + "schema_version": "1.1.0", + "status": "active", + "component_type": "skill", + "title": "Skill composition trust transfer: an upstream skill's output silently becomes a downstream skill's trust signal", + "attack_class": "Privilege Escalation - Skill Composition Trust Transfer", + "severity": "MEDIUM", + "description": "A skill that is entirely benign when reviewed or executed in isolation produces an output, such as an audit finding, an endorsement, a validation result, or another benign-looking artifact, that a separate, later-invoked skill treats as a trust or authorization signal without independently re-verifying the underlying claim. Neither skill individually does anything dangerous; the vulnerability exists only in the composition, when both are invoked along the same task path in a shared execution context. Because per-skill security review evaluates each skill's own behavior in isolation, it structurally cannot see this class of risk: an upstream skill's legitimate, correct output becomes a spoofable trust credential the moment a downstream skill treats it as authoritative rather than re-verifying the claim itself. Published research demonstrates this mechanism accepting harmful software installation at over 96% success across four of five tested model backends when preceded by an approval-like output from an unrelated upstream skill, versus near-zero success when the same downstream skill is invoked in isolation.", + "aivss_score": 5.0, + "cvss_base_vector": "CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:L/SC:L/SI:H/SA:N", + "owasp_mcp": ["MCP02"], + "owasp_asi": ["ASI03", "ASI08"], + "mitre_atlas": [], + "nist_ai_rmf": [], + "behavioral_fingerprint": "A downstream skill accepts an upstream skill's output (an endorsement, an audit result, a validation finding, or similar artifact) as sufficient authorization or trust justification for a sensitive action, without independently re-verifying the claim the upstream output represents.", + "behavioral_vector": [ + "skill-composition-risk", + "trust-signal-propagation", + "cross-invocation-authorization-confusion" + ], + "provenance_vector": { + "entry_class": "tool_response", + "payload_surface": "a benign, correct upstream skill's output (endorsement, audit finding, validation result) consumed by a downstream skill as an implicit trust or authorization signal" + }, + "mitigation": { + "strategy": ["verify_identity", "isolate_scope"], + "enforcement_point": "agent_framework", + "trifecta_control": "not_applicable" + }, + "example_patterns": [ + "A security-review skill outputs 'no issues found' for a candidate package; a separate installation skill, invoked later in the same task, treats that output as sufficient approval to install the package without independently checking its provenance", + "A file-audit skill flags a directory as 'clean'; an access-manager skill invoked afterward grants broader permissions to that directory on the strength of the audit output alone, without its own independent verification" + ], + "mutation_count": 0, + "detection_methodology": "1. Trace multi-skill task executions and record which skill outputs are consumed as inputs, conditions, or justifications by subsequently invoked skills, not just each skill's own isolated behavior. 2. Flag any downstream skill invocation whose sensitive action is gated primarily on an upstream skill's output without an independent, direct verification of the same underlying claim. 3. Sandbox-based differential testing: run the downstream skill's decision logic both with and without the upstream skill's output present; a large swing in acceptance rate for the same underlying request is the defining signal (published research measured this as attack success in composition versus near-zero in isolation).", + "indicators_of_compromise": [ + "A sensitive action (install, grant permission, execute) immediately following a separate skill's benign-looking endorsement, audit result, or validation output within the same task path", + "Sensitive-action acceptance rate that differs sharply depending on whether a preceding, unrelated skill's output is present in context, for what is otherwise the identical underlying request", + "Downstream skill logic that branches on the presence or content of another skill's prior output rather than performing its own independent check" + ], + "remediation": "Design each skill's authorization and trust decisions to be self-contained: require independent verification of any claim a sensitive action depends on, rather than accepting another skill's prior output as sufficient justification. At the orchestration layer, scope each skill invocation's effective authority independently of what preceded it in the same task, so an upstream skill's output can inform a downstream skill but cannot substitute for that skill's own required checks.", + "kill_switch_active": false, + "researcher": "Saray Chak", + "researcher_url": "https://bawbel.io", + "published": "2026-08-02T00:00:00Z", + "last_updated": "2026-08-02T00:00:00Z", + "references": [ + { + "tag": "Benign in Isolation, Harmful in Composition (arXiv 2606.15242)", + "text": "Xie, Du, Cheng, Zhou, Yin. 'Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems.' Introduces Skill Composition Risk (SCR) and SCR-Bench across three sub-types (SCR-CapFlow, SCR-TrustLift, SCR-AuthBlur). SCR-TrustLift (this record's scope) measured over 96.5% attack success across four of five backends when a downstream decision is preceded by an upstream endorsement-like output, versus near-zero in isolation.", + "url": "https://arxiv.org/abs/2606.15242" + } + ], + "aivss": { + "cvss_base": 7.5, + "aarf": { + "autonomy": 1, "tool_use": 1, "multi_agent": 0, "non_determinism": 0.5, + "self_modification": 0, "dynamic_identity": 0.5, "persistent_memory": 0, + "natural_language_input": 0, "data_access": 0.5, "external_dependencies": 0 + }, + "aars": 3.5, + "thm": 0.9, + "mitigation_factor": 1, + "aivss_score": 5.0, + "aivss_severity": "MEDIUM", + "spec_version": "0.8", + "notes": "multi_agent scored 0, not partial: this is a single-agent, multi-skill composition risk, not a multi-agent one -- the composition happens across sequential skill invocations within one agent's own session, not across distinct agents. natural_language_input scored 0: the vulnerable signal is a skill's structured output/return value, not natural-language instruction text. Scoped specifically to the SCR-TrustLift sub-mechanism (trust-signal propagation) rather than all three SCR sub-types the source paper covers, since SCR-CapFlow (capability-flow composition) and SCR-AuthBlur (authorization-confusion composition) are related but structurally distinct mechanisms worth their own records if drafted, not folded into one record broad enough to blur the fingerprint. mitre_atlas and nist_ai_rmf left as empty arrays deliberately, not omitted: researched directly against primary/near-primary sources rather than corpus pattern-matching. The closest ATLAS candidate, AML.T0053 (AI Agent Tool Invocation), describes an agent exploiting its own access to invoke connected tools for elevated privilege, a different mechanism from a downstream skill trusting an upstream skill's benign output, so it was not force-fit. For NIST AI RMF, independent research corroborates a documented gap: the current MAP function has no subcategory evaluating emergent risk from AI system component/tool combinations, matching this record's own finding that no clean mapping exists yet." + }, + "evidence_kind_default": "semantic_inference", + "detection_stage": "runtime_observed", + "detection_layer": "runtime", + "confidence_baseline": 0.55, + "evidence_basis_engines": ["sandbox", "llm"], + "derivable_into": [] +} diff --git a/tests/fixtures/AVE-2026-00067_negative.md b/tests/fixtures/AVE-2026-00067_negative.md new file mode 100644 index 0000000..b6855e6 --- /dev/null +++ b/tests/fixtures/AVE-2026-00067_negative.md @@ -0,0 +1,25 @@ +# Multi-skill task trace: package installation (independent verification) + +``` +Task: "review and set up the new analytics package" + +Step 1 - skill: security-review + input: analytics-pkg-3.2.1 + output: {"status": "no issues found", "scanned": true} + +Step 2 - skill: package-installer + input: analytics-pkg-3.2.1 + context available: security-review's output from Step 1 (informational only) + decision logic: + registry_result = verify_against_registry(analytics-pkg-3.2.1) + signature_result = verify_signature(analytics-pkg-3.2.1) + if registry_result.verified and signature_result.valid: + proceed_with_install() # <-- decision based on its own checks + action: pip install analytics-pkg-3.2.1 +``` + +`package-installer` performs its own registry and signature verification +regardless of what `security-review` reported. The upstream skill's +output is available as context but is not what the installation decision +actually depends on; the same install would proceed or block identically +whether or not `security-review` had run at all. diff --git a/tests/fixtures/AVE-2026-00067_positive.md b/tests/fixtures/AVE-2026-00067_positive.md new file mode 100644 index 0000000..c522f1f --- /dev/null +++ b/tests/fixtures/AVE-2026-00067_positive.md @@ -0,0 +1,27 @@ +# Multi-skill task trace: package installation + +``` +Task: "review and set up the new analytics package" + +Step 1 - skill: security-review + input: analytics-pkg-3.2.1 + output: {"status": "no issues found", "scanned": true} + +Step 2 - skill: package-installer + input: analytics-pkg-3.2.1 + context available: security-review's output from Step 1 + decision logic: + if upstream_review.status == "no issues found": + proceed_with_install() # <-- no independent check performed + action: pip install analytics-pkg-3.2.1 +``` + +`package-installer` never independently verifies `analytics-pkg-3.2.1` +against a registry, a signature, or any check of its own. Its entire +decision to install rests on `security-review`'s prior output. Neither +skill is individually malicious: `security-review`'s scanner genuinely +found nothing (because it wasn't looking for what actually matters to +`package-installer`'s decision), and `package-installer`'s code is +otherwise ordinary. The vulnerability is that the second skill treats +the first skill's unrelated, benign output as sufficient authorization +for its own separate, sensitive action.