Purplelink Daily Digest #38 — July 30, 2026

Matthew Green's technical read on the Anthropic HAWK result provides the cryptographic context that the press coverage omits, specifically what class of attack was found and why it evaded prior analysis. Green is one of the few people positioned to assess whether this represents a genuine capability leap in AI-assisted cryptanalysis or a narrower result that happened to hit a vulnerable target. Connects to: Mythos attack on 3rd-round PQC algorithm candidate puts it out of commission.

Håkon Måløy demonstrated a self-replicating prompt injection worm targeting Copilot for Word: a document containing hidden instructions causes Copilot to embed those instructions into any new document it generates from that source material, propagating the payload without user interaction. The worm mechanism exploits the fact that Copilot treats document content as trusted context, a design assumption that is architecturally difficult to fix without breaking the core use case. Any enterprise using Copilot for Word in document-heavy workflows, legal, finance, HR, should treat untrusted document ingestion as an active attack surface.

Anthropic's automated vulnerability discovery pipeline is generating Microsoft bug reports faster than Microsoft's patch cadence can absorb them, creating a growing backlog of known-but-unpatched vulnerabilities. The strategic implication is that AI-assisted offensive research has already outpaced the institutional capacity of even the largest software vendor to respond, which inverts the traditional assumption that defenders have time to patch before attackers find issues independently. The open question is whether this backlog is being shared with CISA or held under coordinated disclosure agreements that could expire before patches ship.

TurboFieldfare runs 4-bit Gemma 4 26B-A4B-IT on any M-series Mac in approximately 2 GB of RAM using a Swift and Metal inference engine, a roughly 13x memory reduction from naive 4-bit loading. For a one-person macOS studio, this means a capable 26B-parameter model fits inside the memory budget of a base M2 MacBook Air without competing with the OS or other apps. The architecture-specific Metal kernels are the key differentiator; portability to non-Apple hardware is not a design goal.

Sapphire Sleet's September 2025 hijack of debug and chalk, two of the most-downloaded npm packages, sat publicly attributed only to crypto theft for ten months before Amazon's attribution to North Korea surfaced the full scope: at least 18 downstream packages and a wallet-draining payload with supply chain reach into millions of projects. The ten-month attribution gap is the operationally significant detail, suggesting DPRK is deliberately staging supply chain compromises to look like financially motivated crime. Security teams that dismissed the original incident as a crypto-theft story should re-examine their dependency exposure.

Laundry Bear (Void Blizzard) is deploying a backdoor called OWAReaper via an Exchange OWA zero-day that survives credential rotation, meaning standard incident response playbooks that rely on password resets as a containment step are insufficient against this actor. The same group previously exploited a Zimbra vulnerability, indicating a pattern of targeting webmail infrastructure specifically for persistence rather than initial access. Organizations running on-premises Exchange or hybrid OWA deployments in government, telecom, and finance are the explicit target profile.

A controlled experiment pitting a human against a Claude agent over one week of texting found the AI produced measurably higher "exploitable trust" scores with targets. The non-obvious finding is that the advantage came not from deception sophistication but from consistency and patience, traits that human social engineers degrade under time pressure. This has direct implications for vishing and spear-phishing defense: behavioral anomaly detection tuned for human inconsistency patterns will miss AI-driven campaigns.

Bitcoin's post-quantum migration roadmap is predicated on specific NIST-approved algorithms holding until quantum hardware matures, and the 60-hour HAWK break directly undermines that assumption. The non-obvious angle is that AI-assisted cryptanalysis may compress the timeline for algorithm failures independently of quantum hardware progress, creating a second threat vector the crypto community has not priced in. Connects to: Mythos attack on 3rd-round PQC algorithm candidate puts it out of commission.

Samsung's semiconductor arm posted 89.2 trillion won in operating income for Q2 2026, a 250-fold year-over-year jump driven by HBM and DRAM supply constraints against AI training and inference demand. The investor skepticism despite record profits reflects a specific concern: Chinese commodity DRAM is approaching cost parity on standard DDR5, and if AI capex plateaus, Samsung's pricing power collapses simultaneously from two directions. The talent drain to SK Hynix documented in MIT Technology Review this week adds an execution risk layer that the profit number alone obscures.

SaaStr's AI marketing agent 10K hit Marketo's API rate limits hard enough to make the platform operationally unusable, triggering a migration of 10+ years of data to Salesforce Marketing Cloud. The non-obvious dynamic is that agentic workloads generate API call volumes that expose rate limit ceilings legacy SaaS vendors never designed for, and those vendors are raising prices simultaneously, creating a forcing function for churn that has nothing to do with feature competition. Indie and small-team builders designing SaaS tools should treat agent-compatible API throughput as a first-class product requirement, not a future roadmap item.

The argument here is that SaaS churn is being driven primarily by vendor price increases and support degradation rather than AI substitution, with AI serving as the convenient narrative cover for structural problems that predate LLMs. The operationally specific claim is that vendors raising prices 12-20% annually with no new features are accelerating their own displacement by making the switching cost calculus easier, not harder. Connects to: Your Agents Are About to Start Firing Your Vendors. Ours Fired Marketo.

Anthropic's Mythos system found a fatal structural weakness in HAWK, a NIST PQC third-round candidate, after years of human cryptanalysis missed it. The implication is not just that HAWK is dead but that AI-assisted cryptanalysis can now surface vulnerabilities in formally vetted algorithms faster than the standards process can respond. The open question is whether Mythos-class tools are being applied systematically to the other PQC finalists, or whether HAWK was a target of opportunity.