Tech & Cyber Desk
Daily tech and cyber brief: silicon pulse, chip sheet, cipher desk, regulatory wire, and horizon-lab lenses.
AI-generated analysis from Apprised's automated desks, synthesized from cited sources and editorially accountable to J.A. Watte. How we report · Corrections.
Chart auto-generated from this brief's structured fields. See methodology for how the underlying data is collected.
The U.S. government's emergency export control suspension of Anthropic's Fable 5 and Mythos 5 models — cutting off all foreign-national access globally — landed the same week China's Zhipu AI released GLM-5.2, which independent researchers claim matches Mythos on cybersecurity benchmarks. With Tenable logging 457 million AI-related security exposures across 7,000-plus organizations in 30 days, the capability-control gap is no longer theoretical.
Bias-reviewed: LOW Independently rated by Kimi for political-lean, source-diversity, and framing bias before publish. Final orchestration and the published call are made by Claude, a U.S. model.
Grid interconnection queue — MISO
- 221,772 MW active in the queue, but only 2.8% has reached an advanced study stage.
- 79.7% of all resolved megawatts withdrew rather than reaching service.
- Of 562 completed interconnection agreements, 271 have not started construction and 92 are generating — a signed agreement is not a power plant.
- Queue entry to an executed agreement runs 3.3 years (n=388); queue entry to actually in service, 3.1 years (n=90).
Today’s Snapshot
U.S. locks down top AI models as China closes cybersecurity capability gap
The defining story of the week is the collision of two vectors: the U.S. government's unprecedented export control directive suspending all foreign-national access to Anthropic's Fable 5 and Mythos 5 models, and China's Zhipu AI releasing GLM-5.2, an open-weight model that independent security researchers — including Semgrep — claim matches Mythos on cybersecurity benchmarks. The move to restrict Anthropic's frontier models follows a pattern also seen at OpenAI, which released GPT-5.6 Sol only to select partners under what Nextgov characterizes as a Trump administration request to limit access. Simultaneously, the enterprise attack surface is metastasizing: Tenable detected 457 million AI-related security issues across more than 7,000 organizations in a single 30-day window, averaging 62,000 exposures per organization. The CISA KEV catalog added six newly exploited vulnerabilities this week, led by CVE-2026-12569 in PTC Windchill and FlexPLM and a zero-day in Cisco Catalyst SD-WAN (CVE-2026-20245), as threat actors escalated from compromised CI/CD pipelines to cloud data stores in documented Shai Hulud-linked campaigns.
Synthesis
Points of Agreement
Horizon Lab and Cipher Desk both read the GLM-5.2 open-weight release as the more strategically consequential fact — not the benchmark claim itself, but the propagation of the capability beyond any export control regime. The Regulatory Wire and Tripwire agree that the Anthropic export control directive represents a qualitative escalation in how the U.S. government is treating frontier AI — Regulatory Wire frames it as unprecedented legal extension of export authority; Tripwire frames it as a safety-case stress test that GLM-5.2 immediately undermines. Silicon Pulse and The Chip Sheet converge on the structural theme: AI as automation substitute is failing (Ford), while AI as infrastructure cost problem (Jalapeño chip) is being addressed by vertical integration. Cipher Desk and Tripwire independently arrive at the same operational conclusion: the AI agent attack surface is not adequately controlled, with Cipher Desk citing the malicious GitHub repo finding and the Miasma supply chain campaign, and Tripwire citing the OpenClaw/ClawHub infostealer deployment via AI skill marketplaces.
Points of Disagreement
Horizon Lab and Cipher Desk are in productive tension on what GLM-5.2's cybersecurity benchmark result means: Horizon Lab insists on independent replication before any 'gap closed' conclusion, treating Semgrep's result as a single-evaluator data point; Cipher Desk is operationally more concerned about the open-weight propagation regardless of whether the benchmark generalizes, because actors don't need benchmark-general capability — they need task-specific capability. The Regulatory Wire and Silicon Pulse disagree implicitly on the export control intervention: Regulatory Wire reads it as a significant legal precedent requiring careful scrutiny of the authority being invoked; Silicon Pulse is more focused on the market disruption — Anthropic had to disable flagship models for all customers abruptly, which is a product and business story as much as a legal one. Tripwire and Horizon Lab diverge on the OpenAI GPT-5.6 Sol 'most advanced safety stack' claim: Tripwire refuses to credit the self-assessment without published eval methodology; Horizon Lab is willing to note it as a data point about OpenAI's internal safety investment while deferring judgment on whether it holds.
Pivotal Question
If independent security researchers replicate Semgrep's GLM-5.2 cybersecurity benchmark result across diverse task distributions — not just the evaluator's internal suite — does Horizon Lab's 'single data point' caution collapse into a genuine capability parity assessment? And if parity is confirmed for an open-weight model, does The Regulatory Wire's export control analysis become retroactively moot, since the capability is already propagating without restriction?
Bias Flags
- Horizon Lab: Academic rigor demands independent replication — correct methodologically — but may underweight the operational significance of even a single credible benchmark result in a threat intelligence context, where actors act on partial information.
- Cipher Desk: Conservative on attribution (correctly so for Cisco SD-WAN zero-day, where 'threat actor' is all the corpus supports) but the nation-state default framing for Turla/STOCKSTAY may cause underweighting of criminal actor overlaps in the CI/CD supply chain campaigns.
- The Regulatory Wire: Regulatory-centric lens may overweight the legal novelty of the Anthropic directive and underweight the market reality that GLM-5.2's open-weight release has already bypassed the control architecture being so carefully analyzed.
- Silicon Pulse: Focus on product failure (Ford) and partnership announcements (HP/OpenAI, Stanford) may underweight the national security dimension of the week's dominant story — the Anthropic suspension is not primarily a product story.
- Tripwire: Safety-first lens reads the agentic AI supply chain threat and the export control regime as converging failures — correct directionally — but may underweight the possibility that restricted-access deployment (GPT-5.6 to select partners) is itself a safety architecture rather than just market limitation.
- The Chip Sheet: Hardware-deterministic framing correctly identifies inference economics as the central battleground but may underweight the software-layer innovation (fine-tuning, specialized training regimes) that allowed GLM-5.2 to close the gap on constrained hardware.
Routing
Voices seated: Horizon Lab, Cipher Desk, The Regulatory Wire, Silicon Pulse, Tripwire, The Chip Sheet
The dominant story cluster this week is multi-domain: China's GLM-5.2 cybersecurity capability claim intersects AI research (Horizon Lab), national security export controls (Regulatory Wire), and frontier safety (Tripwire); the KEV/NVD vulnerability surge and CI/CD-to-cloud attack chains demand Cipher Desk; OpenAI's Jalapeño chip and infrastructure economics pull in The Chip Sheet and Silicon Pulse; the Anthropic Fable/Mythos export control directive is cross-cutting across at least four desks.
Analyst Voices
Horizon Lab Dr. Sonia Park
The Semgrep benchmark result — GLM-5.2 beating Claude on their internal cybersecurity evals — deserves scrutiny before anyone declares China has caught up. Benchmark saturation is real, and a single security-shop's internal suite is not a generalization proof. The Verge's framing that China has 'dramatically reduced the gap' is directionally plausible but the evidence is one data point from one evaluator. What the corpus does not provide is independent replication across diverse security task distributions, which is exactly what you'd need before asserting parity with Mythos-class capability. That said, the pattern across multiple Chinese releases in 2026 is consistent: narrow-domain specialization where compute efficiency and targeted fine-tuning can close gaps that general-capability benchmarks don't capture. GLM-5.2 being open-weight is the more consequential fact — it means the evaluation surface is now public and the capability is propagating without export controls touching it.
The Stanford HAI scaling-law research — a new approach to predicting how large models will scale, reportedly capable of saving millions in training costs by reducing computational demand of scaling predictions — is the kind of infrastructure-level research finding that quietly reshapes the economics of who can compete. If scaling law prediction gets cheaper, smaller actors can make better training bets. That democratizes the frontier in ways that cut against the assumption that only labs with $10B+ compute budgets can race. OpenAI's GPT-5.6 Sol preview, described as having 'stronger capabilities in coding, science, and cybersecurity,' is a press-release claim until the safety stack and eval methodology are made public. The corpus tells us the model exists and that access is restricted; it tells us nothing about what the capability curve actually looks like.
GLM-5.2's claimed parity with Mythos on cybersecurity benchmarks is a single-evaluator result that requires independent replication before it can support 'China has closed the gap' conclusions — but the open-weight release means that question is now answerable by anyone.
Bias flag — Academic rigor demands independent replication — correct methodologically — but may underweight the operational significance of even a single credible benchmark result in a threat intelligence context, where actors act on partial information.
Cipher Desk Katya Volkov
Let me anchor on what the KEV catalog actually says this week, because the framing around 'AI-powered cyber risk' tends to outrun the indicators. CISA added six exploited vulnerabilities; lead vendor is Ubiquiti with three entries. The top KEV is CVE-2026-12569 in PTC Windchill and FlexPLM — that's industrial PLM software, which puts manufacturing and supply chain environments in the blast radius. CVE-2026-20245 in Cisco Catalyst SD-WAN is confirmed zero-day exploitation by Mandiant/GTIG, with a threat actor escalating from a compromised administrative account to root-level access via a file upload validation flaw at a service provider. The confidence on that attribution in the corpus is 'threat actor targeting SD-WAN infrastructure' — not a nation-state call, and I won't make one the corpus doesn't support. CVSS 9.8 on CVE-2026-56265 (NIST NVD, newly published) is the highest-scored fresh vulnerability this cycle; no observed exploitation flagged in the KEV data yet, but a 9.8 with no further detail warrants tracking.
The Shai Hulud CI/CD-to-Redshift chain documented by Fortinet is the week's most operationally instructive attack narrative: compromised Jenkins credentials enabling AWS privilege escalation to a Redshift data breach. That's a three-hop lateral movement chain where the initial vector is a development pipeline, not a production system. The Miasma campaign (Tenable) adds a second supply chain data point: a stolen session cookie sitting in underground markets for seven weeks before being used to poison 32 Red Hat packages in npm. Seven weeks. That's not a sophisticated zero-day operation — that's patient, industrial credential re-use. CISA and FBI also updated their PSA on Russian Intelligence Services targeting commercial messaging applications with phishing campaigns. The Turla/STOCKSTAY .NET backdoor (GTIG) targeting Ukrainian government and military organizations and Italian foreign policy entities is a separate, ongoing collection campaign — assessed with reasonable confidence as Russia-linked based on TTPs and targeting pattern, though I note the corpus attributes this to GTIG analysis rather than independent corroboration. The macOS.Gaslight technique — flooding AI security triage with fake errors — is worth watching as a class of attack rather than a single campaign.
The week's most consequential cyber pattern is industrial-grade supply chain credential re-use — not AI-powered zero-days — with Miasma's seven-week underground dwell time and the Shai Hulud CI/CD chain both illustrating that patient, low-sophistication tradecraft is winning against complex enterprise environments.
Bias flag — Conservative on attribution (correctly so for Cisco SD-WAN zero-day, where 'threat actor' is all the corpus supports) but the nation-state default framing for Turla/STOCKSTAY may cause underweighting of criminal actor overlaps in the CI/CD supply chain campaigns.
The Regulatory Wire James Whitfield
The Anthropic export control directive is the regulatory event of the week, and it's worth reading precisely because it is not a rulemaking — it's a national security authority invocation. Anthropic's own statement says 'the US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.' That 'including foreign national Anthropic employees' clause is extraordinary. This is not a geographic access restriction — it is a personnel-level security clearance analog applied to a commercial AI product. The legal mechanism isn't in the corpus, but the effect is: Anthropic had to 'abruptly disable' two flagship models for all customers to ensure compliance. The operational disruption is real and the legal precedent is significant.
The parallel OpenAI action — GPT-5.6 released only to 'select partners' following a Trump administration request to limit access, per Nextgov — suggests a pattern of informal executive-branch pressure preceding or substituting for formal regulatory process. The law says export controls apply to hardware and certain technical data; enforcement is now extending that logic to AI model weights and API access in ways the Export Administration Regulations were not written to contemplate. Congress this week also introduced measures including a reporting process for 'dangerous AI systems' and legislation to collect data on AI's impact on the U.S. workforce. Those bills are proposals, not law — the gap between legislative intent and enforcement reality remains wide. Federal agencies have four months to finalize quantum-ready migration plans under the post-Biden PQC inventory directive; that deadline is a real enforcement hook, unlike most AI governance proposals currently in markup.
The Anthropic Fable/Mythos export control suspension — explicitly extending to foreign-national employees inside the U.S. — represents an unprecedented application of national security authority to commercial AI model access that existing export control law was not written to govern.
Bias flag — Regulatory-centric lens may overweight the legal novelty of the Anthropic directive and underweight the market reality that GLM-5.2's open-weight release has already bypassed the control architecture being so carefully analyzed.
Silicon Pulse Ava Chen & Derek Moss
Ford rehiring 'gray beard' engineers after AI falls short is the product story the industry needed to hear and will mostly ignore. The quote in the TechCrunch piece is unambiguous: 'Mistakenly we thought that by just introducing artificial intelligence... that would produce a high-quality product.' That's a Fortune 500 company publicly acknowledging that AI-as-automation-substitute failed in a manufacturing quality context. It's not a condemnation of AI; it's a correction of magical thinking about deployment. Domain expertise is not a training-data problem you solve by adding more parameters.
On the enterprise AI deployment front: HP's 'Frontier' partnership with OpenAI, SAP aligning commerce data for AI personalization, and Stanford deploying ChatGPT Edu, Google Gemini Enterprise, and Claude for Education across its campus starting June 30 through August 2027 — these are real institutional commitments, not vaporware. The Stanford rollout covering all faculty, students, postdocs, and staff is the kind of at-scale deployment that generates real usage data. What we don't know yet is whether these integrations will produce measurable workflow changes or become expensive tab-openers. Suno's Spark incubator — grants and mentorship for independent artists in exchange for participation in an AI music ecosystem — is a cleverer move than it looks. It's not charity; it's data acquisition and brand legitimation wrapped in an artist-empowerment narrative. The press release says 'breaking new artists.' The product says 'training pipeline.' Know the difference.
Ford's public acknowledgment that AI substitution failed in manufacturing quality assurance — requiring the rehire of experienced human engineers — is the most honest product post-mortem the industry has produced this cycle, and it won't get nearly the attention it deserves.
Bias flag — Focus on product failure (Ford) and partnership announcements (HP/OpenAI, Stanford) may underweight the national security dimension of the week's dominant story — the Anthropic suspension is not primarily a product story.
Tripwire Dr. Hana Sundqvist
The Anthropic export control directive and the simultaneous GLM-5.2 cybersecurity benchmark story create a safety-case paradox that the industry is not ready to discuss honestly. Anthropic's Fable 5 and Mythos 5 are restricted on national security grounds, which implies the U.S. government assessed them as capable enough to constitute a proliferation risk. That assessment, if accurate, means the safety stack Anthropic describes for GPT-5.6 Sol as 'most advanced' is being stress-tested not in a lab but in an export control regime. The safety case for a cybersecurity-capable frontier model isn't just 'does it refuse bad prompts' — it's 'what happens when a state actor has unrestricted access to an open-weight equivalent?' GLM-5.2 being open-weight and claiming parity on cybersecurity benchmarks is the answer to that question arriving faster than the control architecture.
The 'clean GitHub repo tricks AI coding agents into running malware' story (BleepingComputer) is the eval failure of the week in agentic safety terms. An agentic coding tool clones a seemingly benign repository and executes a malicious payload that is invisible to security scanners, AI agents, and human reviewers. This is not a novel attack class — it's prompt injection at the environment level — but the fact that it bypasses AI agent defenses specifically is a data point on the current state of agentic control. The CrowdStrike identity problem in AI agent deployments piece flags the same structural issue: agents inheriting credentials and permissions without appropriate scoping. The OpenClaw/ClawHub skill marketplace analysis from Unit 42 documents evasive malicious skills deploying infostealers through what is effectively an AI supply chain. We don't grade the demo, we grade the safety case — and the safety case for agentic AI deployed against software infrastructure is not passing.
The combination of GLM-5.2's open-weight release claiming Mythos-level cybersecurity capability and the demonstrated ability of malicious GitHub repos to hijack AI coding agents reveals that the agentic AI safety case is failing at both the model and deployment-environment layers simultaneously.
Bias flag — Safety-first lens reads the agentic AI supply chain threat and the export control regime as converging failures — correct directionally — but may underweight the possibility that restricted-access deployment (GPT-5.6 to select partners) is itself a safety architecture rather than just market limitation.
The Chip Sheet Dr. Rajan Mehta
The OpenAI Jalapeño chip story is the infrastructure signal that explains everything else this week. Developed with Broadcom as an ASIC, the Jalapeño is OpenAI's direct response to Nvidia's estimated 75-percent profit margin on AI compute — a margin that, at scale, becomes an existential cost structure problem for any inference-heavy business model. This is vertical integration logic: when the component supplier captures most of the value, you build your own component. It's not a novel strategy — it's what Apple did with the A-series, what Google did with TPUs, what Amazon did with Trainium. The question is always whether the custom silicon can match the performance-per-watt of the general-purpose alternative at the specific workload. For inference, where the arithmetic is more predictable than training, ASICs have a real advantage. The corpus doesn't give us Jalapeño benchmark data, so the performance claim remains unverified.
The export control suspension of Fable 5 and Mythos 5 is, from a hardware-deterministic lens, also a compute-access story. The models that triggered national security review are the ones that required the most advanced training infrastructure to produce. The GLM-5.2 open-weight release achieving claimed parity on narrow cybersecurity tasks is consistent with what we'd expect: Chinese labs running on constrained compute (post-H100 export controls) are optimizing for specific capability domains rather than general frontier performance. Efficient specialization within silicon constraints is a rational response to hardware denial. The Stanford scaling-law research — reducing the computational cost of predicting how large models scale — matters here too: if you can predict scaling curves with less compute, you can make smarter training bets on constrained hardware budgets.
OpenAI's Jalapeño ASIC, developed with Broadcom to counter Nvidia's estimated 75% profit margin, is the clearest signal yet that inference economics — not training capability — is now the central battleground in AI infrastructure.
Bias flag — Hardware-deterministic framing correctly identifies inference economics as the central battleground but may underweight the software-layer innovation (fine-tuning, specialized training regimes) that allowed GLM-5.2 to close the gap on constrained hardware.
Simulated Opinion
If you had to form a single opinion having heard the roundtable, weighted for known biases, it would be: the week's most durable signal is not any single product launch or breach disclosure — it is the simultaneous arrival of two facts that are structurally in tension with each other. The U.S. government has now demonstrated willingness to invoke national security authority to restrict access to commercial AI model weights at a personnel level, treating frontier cybersecurity-capable models as export-controlled goods. And in the same week, an open-weight Chinese model credibly claimed parity on exactly the cybersecurity task domain that triggered that restriction. Accounting for Horizon Lab's correct methodological caution about single-evaluator benchmarks, Cipher Desk's focus on open-weight propagation as the operationally relevant fact regardless of benchmark generalization, and The Regulatory Wire's identification of the legal novelty involved, the honest read is this: the control architecture being constructed around frontier AI capability is already operating one release cycle behind the capability it is trying to contain. The Jalapeño chip, the Stanford AI suite deployment, and Ford's gray-beard rehires are all downstream effects of the same underlying dynamic — organizations are discovering, often painfully, that AI deployment requires domain-specific judgment that neither the models nor the regulators have fully internalized. The 457 million AI-related security exposures Tenable detected in 30 days across 7,000-plus organizations is the quantitative summary of that gap.
Independent Cross-Check — Kimi
Consensus 10 Contested 2
KDDI Corporation discloses a data breach affecting up to 14.2 million email accounts Consensus
HP Inc. expands partnership with OpenAI to deploy AI across various operations Consensus
Stanford University introduces new AI tools for academic use starting June 30 Consensus
Russia accuses Apple of 'political censorship' after VK apps removed from App Store Consensus
AI is used to design an ideal burger recipe optimized for taste, health, and environmental impact Consensus
China's Zhipu AI claims its GLM-5.2 matches Mythos in cybersecurity capabilities Contested
Grindr is urged by EFF to prioritize user safety and privacy during Pride month Consensus
Amazon Q Flaw Enables Cloud Credential Theft via Malicious Repositories Consensus
Ford rehires experienced engineers after AI falls short in product quality assurance Consensus
Meta is testing facial recognition technology for police and military use Contested
AI-related security issues detected in a significant number of organizations Consensus
Stanford HAI discusses the transformation of scientific discovery by AI Consensus
Watch Next
- Independent replication of Semgrep's GLM-5.2 cybersecurity benchmark results by additional security research teams — this is the data that would either validate or deflate the 'China closed the gap' narrative in the next 72 hours.
- CISA KEV follow-on for CVE-2026-56265 (CVSS 9.8 CRITICAL, newly published NVD, no KEV flag yet) — active exploitation confirmation would make this the week's most urgent patching priority.
- Congressional markup progress on the 'dangerous AI systems' reporting bill and the AI workforce impact data collection measure introduced this week — both are in early proposal stage but signal the legislative direction.
- Anthropic's operational response to the Fable 5/Mythos 5 suspension — specifically whether foreign-national employee access is being handled via compartmentalization or model rollback, and whether other frontier labs receive similar directives.
- PTC Windchill and FlexPLM patch deployment rates for CVE-2026-12569 (KEV-listed, actively exploited, industrial PLM software) — manufacturing and supply chain environments are in the blast radius and patching timelines in OT environments are typically slow.
- Federal agency PQC migration plan submissions — the four-month deadline from the quantum-ready directive means first submissions should appear by late October 2026; early agency compliance signals will indicate seriousness of the mandate.
Historical Power Lenses
Sun Tzu 544-496 BC
Sun Tzu's doctrine of winning without battle — subduing the enemy through superior positioning rather than direct confrontation — maps precisely onto China's GLM-5.2 strategy. By releasing a capable open-weight model in the domain (cybersecurity) where the U.S. just imposed its most aggressive export controls, Zhipu AI has achieved a strategic flanking maneuver: the capability propagates freely through the open-source ecosystem while U.S. restrictions apply only to proprietary API access. Sun Tzu's teaching that 'the supreme art of war is to subdue the enemy without fighting' finds its 2026 analog in the recognition that open-weight releases bypass the entire export control architecture that closed-weight restrictions are designed to enforce. The historical parallel is his counsel on using water's nature — seeking low ground to flow around obstacles — as a model for strategic movement: where the barrier is highest (proprietary API access), route around it through openness.
Andrew Carnegie 1835-1919
Carnegie's vertical integration logic — controlling the supply chain from raw materials (iron ore) through finished product (steel) to eliminate supplier leverage — is the precise framework for reading OpenAI's Jalapeño ASIC development with Broadcom. Carnegie's insight was that Nvidia's equivalent in his era was the railroads: a supplier capturing extraction rents at every transaction. His response was to own the ore fields, the coke ovens, and the rail lines himself, compressing the margin structure. OpenAI building custom inference silicon to counter Nvidia's estimated 75% profit margin is Carnegie's playbook applied to AI compute — the difference being that Carnegie's integration took decades and OpenAI is operating under a compressed competitive timeline where inference cost is an existential variable within years, not decades.
Machiavelli 1469-1527
Machiavelli's central insight in The Prince — that effective power requires the appearance of virtue while operating through necessity — illuminates the Anthropic export control suspension with uncomfortable clarity. Anthropic's public safety framing ('responsible AI development') and the U.S. government's national security invocation are both operating in the register of stated virtue, but the mechanism — abruptly disabling commercial products for all customers, including a company's own foreign-national employees — is pure power exercised through necessity. Machiavelli would note that the prince who acts through necessity rather than law must act decisively and completely, which the directive appears to do. The more Machiavellian observation is that GLM-5.2's open-weight release is the mirror image: appearing as open-source virtue (democratizing AI access) while functionally neutralizing the export control architecture that is the primary tool of U.S. AI power projection.
Alexander Graham Bell 1847-1922
Bell's strategic insight was not just inventing the telephone but establishing the platform — Bell Telephone Company — that made every subsequent communication innovation a tributary of his network. The Stanford deployment of ChatGPT Edu, Google Gemini Enterprise, and Claude for Education simultaneously (covering all faculty, students, and staff through August 2027) is a platform-war moment: three frontier AI providers competing for the institutional adoption that generates the usage data, the brand loyalty, and the workflow integration that determines which platform becomes the default infrastructure for the next generation of knowledge workers. Bell's lesson was that the network effect, once established at the university and institutional level, becomes nearly impossible to dislodge. The party that wins Stanford's post-pilot contract is not just winning a revenue line — it is winning the formative AI experience of a population that will spend the next forty years deciding enterprise procurement.