Purplelink Daily Digest #4 — June 25, 2026

Anthropic alleges Alibaba used API access to systematically extract Claude's capabilities into a competing model, a form of model distillation attack via commercial terms violation rather than a technical exploit. This is the first major public accusation of capability extraction at this scale between a US frontier lab and a Chinese hyperscaler, and it sets a precedent for how model providers will need to instrument API usage for behavioral fingerprinting. The legal and technical question of what constitutes 'illicit extraction' versus legitimate fine-tuning on model outputs remains entirely unresolved.

OpenAI's Jalapeno ASIC is purpose-built for LLM inference rather than training, signaling that OpenAI's primary cost pressure is now serving queries at scale rather than running gradient updates. Designing around inference workloads means optimizing for memory bandwidth and low-latency token generation rather than FLOPS, which is a fundamentally different architectural target than NVIDIA's H100/H200 line. The Broadcom partnership mirrors Google's TPU strategy and puts direct pressure on NVIDIA's inference revenue, which is increasingly where the margin lives.

Google shipping computer-use capability in Gemini 3.5 Flash, a cost-optimized model, suggests the feature is moving from frontier-model novelty to commodity infrastructure faster than expected. Putting agentic screen-control in a Flash-tier model means the cost-per-action drops dramatically, which changes the threat model for automated social engineering and credential harvesting at scale. Security teams building detection for AI-driven browser automation need to account for this capability being accessible at near-zero marginal cost.

CVE-2026-20245 (CVSS 7.8) in Cisco Catalyst SD-WAN was exploited in the wild at least two months before public disclosure, with attackers using rogue SD-WAN peering to gain admin and root-level access without valid credentials. The rogue peering vector is non-obvious: it abuses the trust model of SD-WAN overlay networks rather than a traditional memory corruption or injection flaw, meaning standard vuln-scanning would not have detected active exploitation. Organizations running Catalyst SD-WAN should treat any unexpected peering relationships in logs as a retroactive indicator of compromise going back to at least April 2026.

Dark Reading's coverage adds the detail that rogue peering was the specific mechanism used to connect to victim SD-WAN devices and escalate to root, confirming Mandiant's attribution of a sophisticated, patient threat actor with deep knowledge of SD-WAN trust architectures. The two-month pre-disclosure exploitation window is consistent with nation-state actors who stockpile zero-days in network infrastructure rather than burning them quickly. Connects to: Mandiant reveals how Cisco SD-WAN zero-day attacks gained root access.

Operation Endgame's second phase, involving Bitdefender, Bitsight, ESET, and Microsoft alongside Europol, recovered 27 million stolen credentials from Amadey and StealC infrastructure, targeting the 'assembly line' model where loaders, stealers, and ransomware are sold as modular services. The 27M credential figure is operationally significant because it represents a known-compromised dataset that defenders can cross-reference against their own identity stores before threat actors rotate to new infrastructure. The simultaneous disruption of both a loader (Amadey) and a stealer (StealC) is strategically smarter than single-tool takedowns, since it breaks the supply chain at multiple points.

Five malicious packages on ClawHub, OpenClaw's AI skills marketplace, bypassed security checks while embedding infostealers, establishing that AI agent plugin ecosystems are replicating the npm/PyPI supply chain attack pattern within months of launch. The fact that these packages evaded automated security scanning suggests the review pipeline for AI skill marketplaces is significantly less mature than package registries that have had years of adversarial pressure. Researchers building dark web intelligence pipelines should monitor ClawHub and similar registries as emerging distribution channels for credential-stealing payloads.

Amazon's additional $13B India AI and cloud commitment through 2030 follows Microsoft's $3B and Google's $2B India pledges in the same cycle, making India the primary non-US geography for hyperscaler AI infrastructure buildout. The strategic logic is data sovereignty compliance plus a large English-language developer base, but the practical implication is that AWS, Azure, and GCP inference capacity in South Asia will expand significantly, lowering latency and cost for AI-native startups building in that region. For AI infrastructure investors, the concentration of three hyperscalers simultaneously committing to the same geography within months of each other suggests competitive signaling rather than independent demand analysis.

Adobe pulling back a planned price increase is the first concrete signal that the 2022-2025 SaaS pricing cycle, where vendors raised prices 10-20% annually and bundled AI SKUs as mandatory add-ons, is hitting customer resistance at enterprise scale. For indie and small-studio software builders on Apple platforms, this is a leading indicator that customers are now actively auditing and cutting SaaS spend, which creates both churn risk for subscription products and an opening for lower-cost alternatives. The more specific question is whether Adobe's retreat is AI-substitution pressure from tools like Firefly competitors, or pure price fatigue.

PayPal's 50% conversion lift on 8,000 previously unworked leads per month using Salesforce Agentforce is notable because the baseline was zero human follow-up, meaning the comparison is agent vs. nothing rather than agent vs. human, which inflates the headline number but also reveals a real untapped revenue pool in mid-market lead queues. The operationally interesting detail is that the leads were abandoned not due to low quality but due to capacity constraints, which is a structural problem AI agents solve without requiring headcount approval. Indie SaaS builders with any outbound motion should treat this as a template for automating the long tail of inbound signups that never get a response.