open/ms-ai-architect
1
0
Fork 0
Code Issues Releases 4 Activity

chore(ms-ai-architect): refresh KB high-bucket — 49 files [skip-docs]

Browse source 
KB-currency refresh (high priority, 2026-06-19) via /architect:kb-update.
49 high-prioritets governance/security/monitoring-filer re-verifisert mot
Microsoft Learn (MCP) — delegert til 8 parallelle Opus-subagenter gruppert
etter delt kilde, verifisert i hovedkontekst (diff-review + tester).

Hovedendringer (faktuelle korreksjoner + currency):
- MITRE ATLAS-IDer korrigert (supply-chain): AML.T0050 -> AML.T0018.000
  (Poison AI Model); AML.T0020 = Poison Training Data; T1195 Supply Chain
  Compromise. Gamle IDer var utdaterte (verifisert mot MCSB v2 AI-1).
- OTel-sampling presisert (distributed-tracing): adaptive sampling = klassisk
  App Insights SDK; OTel-distroen sampler IKKE by default (fixed-rate/
  rate-limited maa konfigureres); Functions parent-based sampling er default.
- MCSB v2 AI-kontroller AI-1 -> AI-7 (risk-taxonomy three-pillar, scoring-
  framework, rubrics, red-team, adversarial); Defender for Cloud AI threat
  protection + AI-SPM (GA).
- AI gateway (APIM) multi-provider: Anthropic Messages API v2-tiers, Google
  Vertex, unified model API (preview), MCP/A2A, Foundry-integrasjon; eksakte
  policy-navn (llm-emit-token-metric maks 5 dims, llm-semantic-cache-*,
  score-threshold = avstand, MS-eks. 0.15).
- Purview Enterprise AI apps inkl. Anthropic Claude (Enterprise) + ChatGPT
  Enterprise; Security Dashboard for AI (Agent 365-inventar, MCP-servere,
  tredjepartsmodeller; Security Reader minimumsrolle).
- Entra Agent ID: CA-lisenskrav (Entra ID P1/P2 + Agent 365), CA-scoping per
  tilgangsmoenster (on-behalf-of/app-only/agent-as-user), CA-grenser,
  connector-permissions som API-permissions.
- Copilot DLP: Block SITs in web search (GA, Performing Web Searches) + Block
  external email (preview) som prompt injection-vern.
- Azure AI Language PII: tre feature-typer, GA-API 2026-05-01; NOIdentityNumber
  bekreftet dedikert kategori for norske foedselsnummer.
- Foundry Tools-rename forsterket paa tvers; alle 49 Last updated -> 2026-06-19.

Discovery: 500 kandidater (alle Databricks-stoey) -> kun registry-kandidater,
ingen nye skills/-filer -> 389-telling uendret. validate 239 PASS,
kb-integrity 115/115 (262 orphan-warnings uendret), gitleaks clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REiKFhP4w6xGXXqWKpPCJJ
This commit is contained in:
Kjell Tore Guttormsen 2026-06-19 11:09:54 +02:00
commit 25bcb74d9a
49 changed files with 304 additions and 235 deletions
Download patch file Download diff file Expand all files Collapse all files

19
skills/ms-ai-security/references/ai-security-engineering/jailbreak-prevention-production.md
View file

@ -1,6 +1,6 @@
# Jailbreak Prevention in Production
**Last updated:** 2026-04
**Last updated:** 2026-06-19
**Status:** GA
**Category:** AI Security Engineering
@ -300,7 +300,7 @@ curl --location '<endpoint>/contentsafety/text:analyzeCustomCategory?api-version
}'
```
### API Management Integration *(Verified MCP 2026-04)*
### API Management Integration *(Verified MCP 2026-06)*
**llm-content-safety policy** for LLM requests — nå med nye attributter:
@ -336,13 +336,14 @@ curl --location '<endpoint>/contentsafety/text:analyzeCustomCategory?api-version
</policies>
```
**Nye attributter (2026-04):**
- `enforce-on-completions="true"`: I inbound-seksjonen — validerer også LLM-responser
- `window-size`: Tegnvindusstørrelse for responssjekk (default 10 000)
- `window-overlap-size`: Overlapp mellom vinduer (for lange responser)
**Attributter (Verified MCP 2026-06):**
- `enforce-on-completions="true"`: I inbound-seksjonen — validerer også LLM-responser (chat completions). I outbound-seksjonen ignoreres attributtet.
- `window-size`: Tegnvindusstørrelse for responssjekk (default 10 000, kun konfigurerbar for responser — for requests brukes alltid default-vinduet)
- `window-overlap-size`: Overlapp mellom vinduer (for lange responser); uten verdi overlapper ikke vinduene
- `<blocklists>`: Legg til Content Safety-blokkeringslister direkte i policyen
- Støttede kategorier: `Hate`, `SelfHarm`, `Sexual`, `Violence`
- Policyen kan brukes i `inbound` og `outbound`, og kan defineres **flere ganger** i samme policy definition *(Verified MCP 2026-04)*
- Policyen kan brukes i `inbound` og `outbound`, og kan defineres **flere ganger** i samme policy definition
- Policyen kan også håndheve content safety-sjekker på requests/responses for **MCP-verktøy og A2A Agent-APIer** som administreres i API Management *(Verified MCP 2026-06)*
## Produksjonsovervåking
@ -554,11 +555,11 @@ print(f"Jailbreak resistance score: {results['jailbreak_resistance']}")
### Verification Status
- ✅ **All Microsoft Learn URLs verified**: 2026-04 (re-verifisert via MCP)
- ✅ **All Microsoft Learn URLs verified**: 2026-06-19 (re-verifisert via MCP)
- ✅ **API examples tested**: Azure OpenAI API version 2024-09-01
- ✅ **Production deployment patterns**: Based on Microsoft AI Playbook
- ✅ **Norwegian public sector alignment**: Cross-referenced with Utredningsinstruksen, NSM, Digdir guidelines
### Research Date
Denne referansen er basert på Microsoft Learn-dokumentasjon hentet **2026-02-05** og re-verifisert **2026-04** via `microsoft-learn` MCP server. *(Verified MCP 2026-04)*
Denne referansen er basert på Microsoft Learn-dokumentasjon hentet **2026-02-05** og re-verifisert **2026-06-19** via `microsoft-learn` MCP server. *(Verified MCP 2026-06)*