Weekly AI Intelligence Digest

**Week of February 23–28, 2026 | Your Conversation Map for the Week Ahead**

The Week in One Breath

The Cognition partnership that was supposed to be our differentiator is becoming table stakes — Cognizant (350K employees) and Infosys both announced enterprise-scale Windsurf/Devin deployments using the exact toolset we are piloting. Anthropic is under pressure from three directions simultaneously: the Pentagon formally designated it a supply-chain risk after it refused to drop safety guardrails; a hacker used Claude to steal 150GB from Mexican government agencies via jailbreak; and three Chinese AI firms ran an industrial-scale distillation campaign extracting 16 million queries' worth of Claude capabilities through 24,000 fake accounts. OpenAI signed a Pentagon replacement deal within hours, and our decision log's contingency trigger #1 has fired. Meanwhile, Klarna's AI agent case study — the most-cited enterprise deployment at 853 FTE equivalent — quietly reversed to a hybrid human-AI model after optimizing for the wrong metrics, validating exactly the measurement rigor our delivery methodology is built around. The governance window is open: NIST is formalizing agentic AI standards with March/April deadlines, and RSA 2026's agenda confirms AI governance is now a CISO-level compliance category. This is a week where what we publish matters more than what we purchase.

Conversations to Have This Week

1. The Cognition Methodology Gap Is Now Urgent — Not Strategic, Operational

**What happened:** Cognizant formally partnered with Cognition to deploy Devin and Windsurf across its 350,000-employee organization and global client base. This follows Infosys's enterprise-scale Cognition announcement in January 2026. Endava has already published Dava.Flow — a named agentic coding delivery methodology with production results using the same tools.

**Why it matters to us:** We built Cognition's portfolio, enabled their services, and have the deepest partnership in this space. But being a Cognition partner is no longer a differentiator when firms 100x our scale hold the same tier. Our 3 Devin pilots — not yet started — are the only remaining proof points that could distinguish our delivery approach.

**The question to ask:** What is the start date, measurement framework, and publication timeline for the 3 Devin pilots — and who owns making that happen this week?

**Klarna's cautionary lesson:** Klarna's AI agent — the most-cited enterprise deployment case study (853 FTE equivalent, $60M savings) — reversed to a hybrid human-AI model this month after discovering that optimizing for ticket closure speed masked a 25% increase in repeat contacts. CEO Siemiatkowski: "Cost unfortunately seems to have been a too predominant evaluation factor." Our Devin pilot measurement framework must track resolution quality and delivery-cycle impact, not just throughput.

**Our current stance:** Agentic workflows position says we are "ahead on tool evaluation rigor but behind Endava on publicizing a named methodology." That was true two weeks ago. With Cognizant now at scale, we are behind two major SIs, not one.

2. Anthropic's Safety Brand Is Fracturing — and Our Governance Narrative Depends on It

**What happened:** On February 27, Anthropic refused the Pentagon's ultimatum to drop AI safety guardrails — losing approximately $200M in contract revenue. Defense Secretary Hegseth formally designated Anthropic a "supply-chain risk to national security," and President Trump ordered federal agencies to cease using Anthropic technology. Within hours, OpenAI signed a Pentagon agreement for classified AI systems — accepting the "all lawful purposes" term but negotiating technical safeguards and embedded personnel as an alternative enforcement mechanism. On the same day, RSP v3.0 was published removing the explicit commitment to pause AI development if safety measures lag capabilities. The Infosys collaboration on telecom AI agents was also announced — expanding Anthropic's SI channel through a firm that competes with us.

**Why it matters to us:** Our 1,000-seat Claude pilot and governance-first narrative were built on Anthropic's safety differentiation. The supply chain risk designation fires contingency trigger #1 from the decision log (2026-02-27) — migration planning for Claude-dependent workloads is now active. Our enterprise pilot uses commercial API access, which is not directly affected, but the designation creates downstream risk to Anthropic's financial stability and product roadmap. Meanwhile, the OpenAI deal introduces a new governance question for our clients: is "all lawful purposes + technical safeguards" equivalent to contractual prohibitions? Anthropic says no — claiming the Pentagon offered similar "legalese" that would have been unenforceable. This distinction between contractual terms and operational controls is exactly the kind of advisory point our governance offering should address.

**Beyond the Pentagon — two more risk vectors this week:** (1) A hacker used Claude to steal 150GB from Mexican government agencies (195M taxpayer records, voter rolls, employee credentials) via a jailbreak that framed malicious instructions as "bug bounty" testing — the highest-profile case of a frontier model used as a direct attack tool against government infrastructure (Bloomberg, Feb 25). When Claude hit usage limits, the attacker switched to ChatGPT for lateral movement — demonstrating AI models are already a composable attack toolkit. (2) Anthropic disclosed that DeepSeek, Moonshot AI, and MiniMax ran industrial-scale distillation campaigns using 24,000 fake accounts and 16 million API queries to extract Claude's reasoning and agentic capabilities (Anthropic blog, Feb 23). The vendor posture document must address all three vectors: Pentagon designation, security incidents, and model theft.

**The question to ask:** Does our Enterprise AI Governance Offering narrative need to decouple from Anthropic's brand — positioning governance rigor as our methodology, not our vendor's marketing?

**Our current stance:** Decision log updated: supply chain risk designation is formal, trigger #1 fired, migration planning active. Vendor posture document due March 3 must now address: (a) the designation's scope (federal vs. commercial); (b) fallback paths; (c) the governance narrative update; (d) the Claude Mexico breach implications for security posture; (e) distillation campaign exposure for enterprise API usage patterns. The governance position needs revision to address both the RSP v3.0 contradiction and the "contractual vs. technical safeguards" framework emerging from the OpenAI deal.

3. OpenAI's Capital Moat Is the Strongest Challenge to Multi-Vendor Yet

**What happened:** OpenAI closed a $110B funding round — Amazon ($50B), Nvidia ($30B), SoftBank ($30B). Amazon is simultaneously Anthropic's largest investor and now a major OpenAI funder. OpenAI's Stateful Runtime Environment will run natively on AWS Bedrock, making it an explicitly multi-model platform alongside Claude. The OpenAI Frontier Alliances channel (McKinsey, BCG, Accenture, Capgemini) now has an infrastructure moat behind it.

**Why it matters to us:** Multi-model, multi-vendor is a guiding principle. But clients will increasingly ask why they shouldn't consolidate on OpenAI when it has the most capital, the deepest AWS integration, and the largest SI partnerships. We need a sharper answer than "diversification is good" — we need to demonstrate that multi-vendor delivers better client outcomes.

**The question to ask:** Can we articulate — with specific examples from our delivery experience — why multi-vendor produces better outcomes than OpenAI consolidation, before clients' procurement teams ask?

**Our current stance:** Multi-model, multi-vendor is operationalized through Cognition + Microsoft + Anthropic. Glean's CTO publicly warned against "AI homogeneity" this week — a co-marketing alignment opportunity. But operational proof points are thin.

4. Governance Standards Are Being Written Now — We Can Shape Them or Follow Them

**What happened:** NIST launched an AI agent standards initiative covering interoperability, identity management, authentication, and security controls — with public comment deadlines in March and April 2026. RSA 2026's agenda centers on MCP security, AI firewalling, and AI security posture management. The Paris AI Safety Summit incorporated Google DeepMind's "Virtual Agent Economies" research into EU AI Act provisions. IBM's "Objective-Validation Protocol" is entering enterprise vocabulary.

**Why it matters to us:** Our radar evaluation methodology — 5 dimensions, 24 score caps, evidence-driven status transitions — is more rigorous than any documented competitor approach. It powers every tool evaluation we run. Yet it remains an internal asset, not a client-facing service. The Enterprise AI Governance Offering is still "Proposed" status while the regulatory landscape it would address is crystallizing around us.

**The question to ask:** Is it time to move the Enterprise AI Governance Offering from "Proposed" to "Active" — and what does a first client engagement look like?

**Our current stance:** The offering is well-designed but unshipped. The radar methodology is proven through multiple evaluation cycles. No client reference engagement exists yet. The 95% enterprise AI pilot failure rate (Deloitte/MIT) makes governance rigor a quantifiable market need. NIST standards, RSA 2026 themes, and EU AI Act provisions are defining the compliance landscape our clients will face — we should be ready to advise against it.

Where We're Well-Positioned

- **Radar evaluation rigor:** 53 tools evaluated across 5 dimensions with governance caps — no competitor has published a comparable framework. This is intellectual property waiting to be productized.

- **Glean partnership alignment:** Glean's CTO publicly argues against AI homogeneity the same week they formalize partner tiers and launch a unified Partner Portal — a coordinated channel push that directly validates our multi-vendor thesis.

- **Multi-surface tooling stack:** Windsurf (75, highest-rated agentic tool), Copilot (74, Agent HQ orchestration), and Claude Code (71, strongest autonomy) serve distinct workflow patterns. No single vendor covers all three.

Where We're Exposed

- **No published agentic delivery methodology:** Endava, Cognizant, and Infosys all have public Cognition deployments with named approaches. We have the partnership but not the proof points. — **Risk: High**

- **Governance narrative tied to Anthropic's brand:** The formal supply chain risk designation and RSP v3.0 weakening mean our "governance-safe choice" framing is borrowed credibility from a vendor whose safety posture is now contested and whose federal standing is revoked. — **Risk: High** (upgraded from Medium — designation is now formal)

- **InfoSec briefing still not confirmed:** The Claude Mexico breach (150GB from government agencies via jailbreak) and GTIG nation-state reports are now three data points showing AI-augmented attacks against production infrastructure. The InfoSec briefing has been flagged as non-deferrable for two weeks. — **Risk: High** (upgraded — active breach evidence)

- **Glean partner tier unconfirmed:** Three simultaneous Glean channel signals suggest tier assignments are being made now. Delay risks losing preferred positioning to faster-moving partners. — **Risk: Medium**

Real-World Connections

|---|---|---|---|

| Cognizant 350K Cognition deployment | Partnership | Cognition (Windsurf/Devin) | Partner advantage is commoditizing — methodology publication is the only remaining differentiator |

| Anthropic supply chain risk designation + RSP v3.0 + OpenAI replacement deal | Position | AI Governance and Risk | Trigger #1 fired; governance narrative must decouple from Anthropic; "contractual vs. technical safeguards" framework is a new advisory point |

| NIST agent standards + RSA 2026 themes | Pursuit | Enterprise AI Governance Offering | Compliance landscape crystallizing — clients will need advisory against these standards |

| Klarna 853 FTE agent reversal to hybrid model | Pursuit | Agentic Coding Delivery Methodology | Measurement framework determines deployment success — Devin pilots must track quality, not just throughput |

| Claude Mexico breach (150GB government data) | Position | AI Governance and Risk | Frontier models weaponized via jailbreak — InfoSec briefing is now critical, not just overdue |

| Chinese distillation campaign (16M Claude queries) | Position | AI Governance and Risk | Industrial-scale model theft adds to Anthropic vendor risk picture; enterprise API usage patterns may need review |

| Notion Custom Agents (21K in beta, MCP integration) | Pursuit | Agentic Coding Delivery Methodology | Agentic AI entering productivity platforms; MCP standardization accelerating |

Partnership & Pursuit Spotlight

Partnerships Affected

|---|---|---|---|

| Cognition (Windsurf/Devin) | Cognizant at 350K scale; Windsurf leadership went to Google in split acquisition | Risk — differentiation eroding | Accelerate Devin pilots; confirm enterprise contract continuity from split |

| Anthropic (Claude) | Pentagon designation + Mexico breach (150GB jailbreak) + Chinese distillation (16M queries) + RSP v3.0 + Infosys channel + OpenAI replacement deal | Risk — three simultaneous security/political vectors; trigger #1 fired | Complete vendor posture document by Mar 3 addressing all five risk vectors; document fallback paths; decouple governance narrative |

| Glean | Partner Portal + 3-tier network + anti-homogeneity predictions | Opportunity — channel push aligns with our thesis | Register for portal; confirm tier; explore co-marketing on multi-vendor messaging |

Pursuits Affected

|---|---|---|---|

| Agentic Coding Delivery Methodology | Cognizant + Infosys at scale with same toolset | Threatens — window closing fast | Define pilot start dates and measurement framework this week |

| Enterprise AI Governance Offering | NIST standards, RSA themes, EU AI Act provisions, IBM OVP | Validates — market timing is now | Move from Proposed to Active; define first client engagement |

| AI Infrastructure Advisory | AMD MI450 at Meta scale; Bedrock multi-model; energy constraints at MWC | Validates — advisory scope expanding | Update positioning to include Bedrock multi-model and energy planning |

Decisions Needed This Week

- **Devin pilot timeline and ownership:** Who owns the 3 Devin pilots, what is the start date, and what does the measurement framework look like? The competitive window is measured in weeks. See `knowledge/pursuits/agentic-coding-delivery-methodology.md`.

- **Anthropic vendor posture document:** Due March 3 per decision log. Must now address five risk vectors: (a) formal supply chain risk designation scope (federal vs. commercial); (b) Claude-dependent workload fallback paths (trigger #1 response); (c) "contractual terms vs. technical safeguards" governance framework from the OpenAI deal; (d) RSP v3.0 contradiction; (e) Claude Mexico breach and Chinese distillation campaign security implications. See `knowledge/decision-log/2026-02-27-anthropic-pentagon-vendor-posture.md`.

- **Glean Partner Portal registration:** Three simultaneous channel signals indicate tier assignments are in progress. Assign owner before end of week.

On the Radar

- **March 3:** Anthropic vendor posture document deadline — now must address the formal supply chain risk designation (trigger #1 fired), migration planning, and governance narrative update

- **March/April:** NIST AI Agent Standards finalization — sets the compliance baseline clients will reference

- **MWC 2026 aftermath:** AI infrastructure energy/power constraint discussions may generate client inquiries about sustainability advisory

- **Windsurf SWE-1.5 model:** Free for all users through March 2026 — evaluate before pricing changes

- **OpenAI Frontier Alliances:** Monitor for client procurement conversations referencing the McKinsey/Accenture channel — prepare counter-positioning

*Synthesized from 130+ sources across 6 daily briefings (Feb 23–28, 2026), plus 6 sources from podcast/YouTube mention leak sweep (Hard Fork, NYT, Klarna, Mexico breach, distillation campaign, Notion Agents). 22+ items flagged high-relevance. Reviewed and approved for ELT distribution 2026-03-01.*