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.
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
AI Agents Go Live, Oracle Zero-Day Burns Higher Ed, Anthropic Models Pulled on Export Order
The week ending June 15, 2026 was defined by a three-front collision: AI systems crossed from research into financial and enterprise execution — Visa integrated ChatGPT into retail purchasing, Coinbase launched agent-driven portfolio trading, and OpenAI committed $150M to a global Partner Network — while the threat landscape sharpened around both legacy and AI-native attack surfaces. CVE-2026-35273, a CVSS 9.8 Oracle PeopleSoft zero-day with active ransomware use, was weaponized by ShinyHunters (UNC6240) against American universities, drawing a CISA KEV listing and an out-of-band patch from Oracle. Microsoft's June Patch Tuesday was the largest on record at 198 CVEs, including three zero-days. Against this backdrop, the Trump administration imposed export controls that forced Anthropic to take its Fable 5 and Mythos 5 models fully offline for all customers — a live test of how sovereign AI policy can instantly reshape the commercial frontier-model market.
Synthesis
Points of Agreement
Silicon Pulse reads the Visa/ChatGPT retail agent and Coinbase for Agents as live financial execution integrations with no settled liability framework; Tripwire independently reaches the same structural conclusion and extends it to the missing safety-case documentation. Cipher Desk and Tripwire agree that the LangGraph SQLi-to-RCE disclosure (Check Point Research) and Unit 42's AI supply chain integrity work identify agentic AI frameworks as a materially under-secured attack surface at the production layer. The Regulatory Wire and Silicon Pulse both identify the Anthropic export control action as a live demonstration that frontier model access is revocable, with direct implications for enterprise partner trust outside the U.S. Horizon Lab and Tripwire share the structural diagnosis that multi-agent safety research — including DeepMind's $10M call — is running behind the deployment cadence, though they differ on how alarming that gap is.
Points of Disagreement
Tripwire and The Regulatory Wire disagree on how to read the Anthropic model takedown: The Regulatory Wire frames it as a landmark enforcement precedent with geopolitical compounding effects; Tripwire insists it is a compliance action, not a safety finding, and objects to the conflation of export control logic with safety-case logic — these are different triggers with different implications for how we should evaluate frontier model governance. Horizon Lab and Tripwire disagree on the severity register for the agentic deployment gap: Horizon Lab notes the research lag as a structural concern to be addressed; Tripwire treats it as an active risk materializing in production systems right now. Silicon Pulse and Cipher Desk hold a productive tension on criminal AI: Silicon Pulse emphasizes the commercial agentic stack being built in public; Cipher Desk emphasizes that criminal actors are using AI as a productivity layer for known attack techniques — the tension is whether the primary risk vector is the legitimate stack being abused or the criminal stack independently maturing.
Pivotal Question
The pivotal question is empirical and time-bounded: if a Visa/ChatGPT retail agent or a Coinbase for Agents deployment experiences a prompt-injection-driven or adversarial-catalogue-driven financial error at scale in the next 90 days, does that event force retroactive safety-case disclosure requirements (moving Tripwire's view toward The Regulatory Wire's enforcement framing), or does it get absorbed as a product incident without triggering the governance response (moving The Regulatory Wire's view toward Silicon Pulse's market-momentum framing)? The answer to that question will determine whether agentic AI safety is a regulatory problem or a product engineering problem.
Bias Flags
- Tripwire: Safety-first lens reads every agentic deployment as a risk materialization; may underweight the operational track record of narrower, better-scoped agent deployments that don't exhibit the blast-radius dynamics of open-ended autonomous systems.
- The Regulatory Wire: Regulatory-centric framing may overweight the Anthropic export control action as precedent-setting when it may reflect a one-off compliance negotiation rather than a durable enforcement framework.
- Cipher Desk: Attribution to UNC6240/ShinyHunters via GTIG is well-sourced here, but the broader framing of criminal AI-as-a-service as 'acceleration of known techniques' may underweight the possibility that the productivity gains are quantitatively significant enough to constitute a qualitative threat-level shift.
- Horizon Lab: Academic rigor around the AIMIP benchmark result and the DeepMind safety funding is well-calibrated, but the dismissal of 'prompted performance' improvements relative to underlying capability may underweight how much the practical deployment envelope matters to enterprise buyers.
- Silicon Pulse: Focus on what is shipping may underweight the liability and failure-mode exposure that Tripwire correctly identifies as absent from the current agentic commercialization push.
Routing
Voices seated: Silicon Pulse, Cipher Desk, The Regulatory Wire, Horizon Lab, Tripwire
This week's corpus clusters around five interlocking domains: AI agent commercialization (OpenAI Partner Network, Visa/ChatGPT, Coinbase for Agents), active cyber exploitation (CVE-2026-35273 Oracle PeopleSoft, ShinyHunters, Microsoft Patch Tuesday), AI model export controls (Anthropic Fable 5/Mythos 5 takedown), AI safety research investment (DeepMind multi-agent funding), and emerging agentic AI risk surfaces (LangGraph SQLi-to-RCE, AI supply chain integrity). Silicon Pulse covers the commercial AI stack build-out; Cipher Desk anchors the Oracle zero-day and criminal AI-as-a-service wave; The Regulatory Wire handles the Anthropic export control action and AI governance; Horizon Lab reads the capability and safety research signals; Tripwire evaluates agentic AI safety cases and the dangerous-capability surface opening under autonomous deployment.
Analyst Voices
Silicon Pulse Ava Chen & Derek Moss
The headline this week isn't any single product — it's the pipeline going live. Visa linking its payment rails directly to ChatGPT so an AI agent can complete a retail checkout without human intervention, Coinbase wiring LLMs into active portfolio trading, and OpenAI dropping $150M into a Partner Network to accelerate enterprise deployment: these are not demos. These are production integrations with real financial execution on the other end. The press release says 'AI-powered commerce.' The product says 'autonomous financial agent with your credit card number.' Know the difference — and know that the liability architecture for when these agents make bad calls hasn't been written yet.
The GitHub trending data underscores where the builder community's head is at. XiaomiMiMo/MiMo-Code landed 8,108 stars in a week. shadcn/improve — a repo specifically designed to have your most capable model audit a codebase and write execution plans for cheaper models — hit 3,777 stars. DietrichGebert/ponytail, explicitly pitched as making your AI agent 'think like the laziest senior dev in the room,' crossed 2,476 stars. The dev community is building the scaffolding for multi-model, cost-optimized agent pipelines. This is the agentic infrastructure layer, assembled in public, at speed.
OpenAI Academy launching three new applied-AI courses is the softer side of the same story: OpenAI is building a workforce development moat alongside its model moat. Preply using OpenAI for personalized language learning feedback is a real deployment, not a concept. The commercialization of the AI stack is happening faster than the enterprise procurement, security, and liability frameworks can absorb it. That gap is where most of the interesting problems — and most of the risk — will live for the next 18 months.
AI agents crossed from demo to financial execution this week — Visa/ChatGPT retail purchasing and Coinbase for Agents are live products, not pilots, with no settled liability framework.
Bias flag — Focus on what is shipping may underweight the liability and failure-mode exposure that Tripwire correctly identifies as absent from the current agentic commercialization push.
Cipher Desk Katya Volkov
CVE-2026-35273 is the anchor for this week's threat picture and it deserves precise treatment. Per CISA's KEV catalog addition on June 12 and Rapid7's emergency threat response, this is a missing-authentication-for-critical-function vulnerability in the Updates Environment Management component of Oracle PeopleSoft Enterprise PeopleTools, CVSS 9.8, remotely exploitable without credentials, with confirmed ransomware use. Oracle issued an out-of-band patch on June 10 — the same day as the advisory — which is the vendor's clearest possible signal about severity. Mandiant and Google Threat Intelligence Group have attributed active exploitation between May 27 and June 9 to UNC6240, the group tracked publicly as ShinyHunters, specifically targeting Oracle PeopleSoft ERP infrastructure at American universities. The higher education sector's disproportionate exposure maps predictably to licensing patterns: large PeopleSoft footprints, historically under-resourced patch cycles, and complex shared-services environments. Attribution to UNC6240 here carries moderate-to-high confidence given GTIG's direct visibility; the ransomware-use flag in the KEV catalog means this is not a theoretical exploitation scenario.
The FBI's disruption of the Chinese phishing-as-a-service operation called Outsider Enterprise — coordinated with Google and Black Lotus Labs, targeting credit card data and passwords across thousands of phishing sites — is a separate but thematically connected data point. The operation demonstrates that criminal AI infrastructure is maturing into service-layer abstraction: a million URLs, industrialized credential harvesting, AI-generated lures. Rapid7's Criminal AI-as-a-Service report released this week makes the same observation: AI in the criminal underground is functioning as a productivity layer for routine operational tasks — phishing content, reconnaissance, social engineering scale — rather than as autonomous hacking systems. That framing matters. The threat is acceleration of known techniques, not emergence of novel attack classes. FortiGuard Labs' analysis of the AsyncRAT campaign using fake AI-themed documents and AutoHotkey loaders confirms the same pattern: AI hype as social engineering surface.
Microsoft's June Patch Tuesday at 198 CVEs — 32 critical, three zero-days with public exploit code — is the largest on record per Krebs and Tenable. The volume alone is a signal about attack surface expansion. The MSNightmare/RoguePlanet repo on GitHub (1,260 stars, C++, described as a Windows Defender vulnerability) warrants a watch flag: a newly starred repo explicitly claiming a Defender bypass is either a security research artifact or a weaponization seed. The indicator confidence on that one is low, but the pattern fits.
CVE-2026-35273, a CVSS 9.8 Oracle PeopleSoft zero-day with confirmed ransomware use, was actively exploited by ShinyHunters (UNC6240) against U.S. universities between May 27 and June 9 — and criminal AI is accelerating known attack techniques, not inventing new ones.
Bias flag — Attribution to UNC6240/ShinyHunters via GTIG is well-sourced here, but the broader framing of criminal AI-as-a-service as 'acceleration of known techniques' may underweight the possibility that the productivity gains are quantitatively significant enough to constitute a qualitative threat-level shift.
The Regulatory Wire James Whitfield
The Anthropic export control action is the most structurally significant regulatory event of the week and it deserves to be read precisely. The Trump administration directed Anthropic to take Fable 5 and Mythos 5 offline to prevent use by foreign nationals, citing national security. Anthropic complied — disabling the models for all customers globally, not just foreign nationals, because access controls at that granularity apparently couldn't be implemented cleanly enough to satisfy the directive. The law says export controls apply to foreign nationals. Enforcement says you pull the model entirely when your identity verification and access architecture can't reliably distinguish. The gap is where the AI export control framework actually operates — and it's a very wide gap. This is not a precedent in the formal legal sense, but it establishes a behavioral template: frontier AI models are now within the practical reach of the export control apparatus, and the compliance burden of differentiating access by nationality is high enough that full suspension becomes the path of least resistance.
The implications compound. If the administration extends this framework to other frontier models — from other labs, including OpenAI, Google, or Meta's open-weight releases — the commercial and geopolitical stakes escalate sharply. OpenAI's $150M Partner Network investment, announced the same week, is now operating under a shadow: enterprise partners in allied but non-U.S. jurisdictions need to understand that frontier model access is a revocable privilege, not a durable commercial commitment. The Regulatory Wire's read is that this action will accelerate EU and allied-nation efforts to develop domestically controlled frontier AI capacity, precisely because dependency on U.S. commercial AI providers has just been demonstrated to be a policy lever, not just a business relationship.
Elsewhere in the regulatory stack: the Senate narrowly rejected Senator Gillibrand's amendment to establish a new Cyber Force service branch, which would have placed it under the Army. The failure is marginal — 'narrowly fails' — suggesting this will return in the 2027 NDAA cycle. California's S.B. 2564, which would ban surveillance pricing, is moving, supported by EFF. The House passed two GOP-backed Treasury fraud-prevention bills that privacy experts, per FedScoop, describe as following the 'DOGE playbook' — broad data aggregation authority with weak privacy guardrails. The MIT CSAIL research finding that AI tools shaping patient care in nearly two-thirds of U.S. hospitals are operating outside regulatory oversight is a slow-building enforcement gap that will eventually force an FDA or HHS response.
The Trump administration's export control directive forcing Anthropic to pull Fable 5 and Mythos 5 globally demonstrates that frontier AI models are now a live instrument of U.S. export control policy — and compliance architecture cannot reliably thread the nationality-access needle.
Bias flag — Regulatory-centric framing may overweight the Anthropic export control action as precedent-setting when it may reflect a one-off compliance negotiation rather than a durable enforcement framework.
Horizon Lab Dr. Sonia Park
Two research signals this week that are worth holding separately. First, Google DeepMind's $10M funding call for multi-agent AI safety research is a meaningful institutional signal — not because $10M is large relative to frontier training budgets, but because it reflects an acknowledgment at the lab level that multi-agent dynamics are under-studied relative to the deployment velocity of agentic systems. The academic community should treat this as a genuine research agenda, not a PR gesture; DeepMind has a credible record of following through on technical safety funding. The question is whether the research cadence can close the gap with the deployment cadence — and this week's Visa/ChatGPT and Coinbase agent integrations suggest the answer is: probably not.
Second, the Allen Institute's AIMIP benchmark release is a genuinely interesting methodological contribution to AI climate science. The finding — that AI climate models can match or beat conventional models on some historical climate metrics while still struggling to generalize reliably to long-term warming trends and unseen climate scenarios — is exactly the kind of benchmark saturation / capability generalization distinction that gets lost in AI-for-science press coverage. The benchmark improved. The generalization to novel climate scenarios did not. The Stanford HAI framing of AI 'simulating 1,000 years of climate in a day' is a real capability, but it should not be read as reliable extrapolation to scenarios outside the training distribution. That distinction matters enormously for policy use of these models.
The Register's piece arguing that 'AI is code and can't be prompted into being smarter' is making a correct but underspecified point. At inference time, yes — a fixed model's capabilities are what they are. But the framing misses the degree to which prompt engineering, retrieval augmentation, and chain-of-thought scaffolding can meaningfully alter effective task performance within a fixed model's weight space. That's not the same as increasing underlying capability, but it's also not irrelevant. The practical performance envelope of a deployed model is wider than its benchmark sheet.
DeepMind's $10M multi-agent safety research call and the AIMIP climate benchmark both flag the same structural gap: deployment of agentic and scientific AI is outrunning the research infrastructure needed to characterize what these systems actually generalize to.
Bias flag — Academic rigor around the AIMIP benchmark result and the DeepMind safety funding is well-calibrated, but the dismissal of 'prompted performance' improvements relative to underlying capability may underweight how much the practical deployment envelope matters to enterprise buyers.
Tripwire Dr. Hana Sundqvist
Three safety-relevant developments this week, and they should not be collapsed into a single narrative. The Anthropic export control takedown of Fable 5 and Mythos 5 is a compliance action, not a safety finding — the models were not pulled because of dangerous capability evaluations but because of national security export controls. These are categorically different triggers and they should not be conflated. Anthropic's Claude Corps fellowship launch and its broader policy posture this week are consistent with a lab trying to demonstrate responsible behavior under regulatory pressure. What we don't have is public safety-case documentation for Fable 5 and Mythos 5 — their capability profiles, their dangerous-capability eval results, their alignment posture. The takedown happened before that information became public. That asymmetry is the problem.
The Check Point Research disclosure of an SQL injection to RCE chain in LangGraph's checkpointer is the agentic safety story that deserves more attention than it's getting. LangGraph is an open-source framework for building stateful multi-agent AI systems with built-in persistence. The checkpointer — the component that stores execution state — was not adequately locked down, enabling an attack chain from SQLi to remote code execution. This is not a theoretical vulnerability in a research demo; it's a production attack surface in a widely-used agentic framework. Unit 42's companion piece on AI agent supply chain integrity — auditing third-party 'skills' for hidden vulnerabilities — maps the same threat surface from the other direction. When you deploy an AI agent that can take actions in the world, every dependency in that agent's stack is a potential attack vector. The safety case for agentic AI deployment cannot be evaluated at the model layer alone.
DeepMind's $10M multi-agent safety research call is a real commitment, but the timeline problem is acute. Horizon Lab is right that the research cadence lags deployment. My framing is sharper: we are deploying agentic AI systems into financial execution, healthcare decision support, and enterprise infrastructure faster than we can characterize their failure modes under adversarial conditions. The Visa/ChatGPT retail agent and Coinbase for Agents are live this week. The safety cases for agentic financial AI — what happens under prompt injection, what happens when the agent encounters a malicious merchant catalogue, what are the blast-radius bounds of an autonomous trading error — are not public. We don't grade the demo. We grade the safety case. The safety case isn't there yet.
LangGraph's SQLi-to-RCE vulnerability in the checkpointer component and Unit 42's AI agent supply chain research reveal that the agentic deployment wave — now live in financial execution via Visa/ChatGPT and Coinbase — has outrun the safety-case infrastructure for its attack surfaces.
Bias flag — Safety-first lens reads every agentic deployment as a risk materialization; may underweight the operational track record of narrower, better-scoped agent deployments that don't exhibit the blast-radius dynamics of open-ended autonomous systems.
Simulated Opinion
If you had to form a single opinion having heard the roundtable, weighted for known biases, it would be: the week of June 15, 2026 marks the moment agentic AI moved from infrastructure to financial execution in a way that can't be walked back — Visa/ChatGPT retail agents and Coinbase for Agents are live, OpenAI's $150M Partner Network is building the enterprise distribution layer, and the developer community (8,108 stars for MiMo-Code, 3,777 for shadcn/improve in a single week) is assembling the multi-model orchestration scaffolding in public. That commercialization wave is real and has momentum. But the safety-case infrastructure for autonomous financial agents, the liability framework for agentic errors, and the security hardening of the frameworks running these agents (per the LangGraph SQLi-to-RCE disclosure) are all materially behind. Simultaneously, CVE-2026-35273's ransomware-linked exploitation of Oracle PeopleSoft at U.S. universities — a CVSS 9.8 zero-day with active KEV listing — and the Anthropic export control takedown both demonstrate that the attack surface and the regulatory surface are expanding faster than the defensive posture. The optimistic read is that these are solvable engineering and governance problems running on a lag. The more cautious read — discounting Tripwire's tendency to see every deployment as a risk and The Regulatory Wire's tendency to see every enforcement action as a durable precedent — is that the lag itself is the risk, and the next 90 days of agentic financial AI in production will be the forcing function that determines whether the industry closes it voluntarily or waits for a consequential failure to force the issue.
Independent Cross-Check — Kimi
Consensus 13
OpenAI launches Partner Network with $150M investment Consensus
FortiGuard Labs analyzes multi-stage malware campaign using fake AI documents Consensus
Google DeepMind announces $10M funding call for multi-agent AI safety research Consensus
FBI dismantles massive AI-powered Chinese phishing service Consensus
ShinyHunters uses Oracle Zero-Day to attack higher education institutions Consensus
Anthropic takes Fable 5 and Mythos 5 offline to comply with U.S. export controls Consensus
Microsoft addresses 198 CVEs in June Patch Tuesday release Consensus
SpaceX goes public with historic IPO Consensus
UC Berkeley receives $13.5 million gift to drive technology to market Consensus
MIT researchers call for regulatory oversight of AI tools in patient care Consensus
Oracle PeopleSoft vulnerability exploited by hackers Consensus
FBI opens Cyber Range in Huntsville for simulating cyberattacks Consensus
Congress seeks controls on AI usage for military targeting and planning Consensus
Watch Next
- Oracle PeopleSoft CVE-2026-35273: Track patch adoption rates at U.S. universities and monitor for additional ShinyHunters (UNC6240) extortion disclosures; CISA BOD 26-04 compliance deadline for federal civilian agencies will be a forcing function.
- Anthropic export control framework: Watch for the administration extending similar export control directives to other frontier labs (OpenAI, Google DeepMind); any clarification on whether access-tier nationality controls could satisfy the directive without full suspension.
- Agentic AI financial execution incidents: Monitor for any reported errors, prompt-injection events, or unauthorized transactions involving Visa/ChatGPT retail agents or Coinbase for Agents in their production deployments.
- MSNightmare/RoguePlanet GitHub repo (1,260 stars, C++, claims Windows Defender vulnerability): Track for weaponization indicators, CVE assignment, or Microsoft security advisory response.
- Microsoft June 2026 Patch Tuesday zero-days (CVE-2026-49160, CVE-2026-50507): With public exploit code available for at least three of the 198 patched CVEs, watch for active exploitation reports in the next 72 hours.
- LangGraph checkpointer SQLi-to-RCE: Track for CVE assignment and patched release from LangChain; enterprise deployments using LangGraph in production should treat this as a critical remediation priority pending a patch.
- Senate Cyber Force amendment: Watch for reintroduction in the 2027 NDAA cycle; the narrow failure signals this has genuine legislative support.
Historical Power Lenses
Andrew Carnegie 1835-1919
Carnegie's defining strategic insight was vertical integration: own the ore, the furnaces, the rail, and the finishing mills, and you control the cost structure of every competitor. OpenAI's simultaneous moves this week — the $150M Partner Network for distribution, the Academy courses for workforce development, the Preply and Visa integrations for application-layer lock-in — are the vertical integration play in AI infrastructure. Carnegie didn't just make steel; he made it impossible to compete with him by controlling every input. OpenAI is attempting the same: model layer, enterprise distribution, developer tooling, and now financial execution rails. The risk Carnegie faced, and that OpenAI now faces, is that vertical integration creates fragility under regulatory attack — Carnegie's empire was eventually forced to restructure, and OpenAI's vertically integrated AI stack is precisely what antitrust and export control regulators are beginning to probe.
Sun Tzu 544-496 BC
Sun Tzu's central asymmetric insight — 'the supreme art of war is to subdue the enemy without fighting' — maps cleanly onto ShinyHunters' campaign against Oracle PeopleSoft. The group did not attack the universities directly; they exploited a missing authentication vulnerability in a back-office ERP component that universities had every reason to believe was lower-priority infrastructure. The attack surface was the unsexy administrative layer, not the research or identity systems the defenders were watching. Sun Tzu wrote about the 'unguarded gate' — CVE-2026-35273 was precisely that, a CVSS 9.8 unguarded gate in a system that processes financial aid, student records, and HR data. The parallel to Sun Tzu's campaign against the state of Wu is apt: he attacked where the enemy was not prepared to defend, not where they were strongest.
Alexander Graham Bell 1847-1922
Bell's lasting strategic achievement was not the telephone itself but the network effect moat he built around it: once enough people were connected to Bell's exchange, the value of being on any competing network collapsed. Visa's integration with ChatGPT this week is an explicit play for the same dynamic in AI-mediated commerce. Visa is not building an AI model; it is inserting its payment rails as the default execution layer for AI agents operating in retail. If AI agents route purchases through Visa's infrastructure by default, the network effect of that positioning compounds with every new agent deployment. Bell understood that the platform — not the terminal device — was the durable asset. Visa is betting that AI agents are the new terminal devices and its payment network is the exchange.
Machiavelli 1469-1527
Machiavelli's counsel in The Prince was that a ruler must be both lion and fox — force when necessary, cunning when sufficient, and never confuse the two. The Trump administration's export control directive forcing Anthropic to pull Fable 5 and Mythos 5 is a lion move: blunt, immediate, demonstrably effective. But Machiavelli would also note the downstream cost: 'He who becomes master of a city accustomed to freedom and does not destroy it, may expect to be destroyed by it.' By demonstrating that U.S. frontier AI access is a revocable privilege, the administration has given every allied nation — and every enterprise outside the U.S. — a concrete reason to fund and develop domestic AI alternatives. The fox move would have been a targeted access control negotiation with Anthropic. The lion move achieved short-term compliance and created long-term strategic incentive for AI sovereignty elsewhere.
Sources Cited
25 sources — show
- Dark Reading
- Rapid7
- Google Cloud / Mandiant GTIG
- CISA
- SecurityWeek
- Nextgov
- Tenable
- Krebs on Security
- Bleeping Computer
- OpenAI
- AI News
- AI News
- Google DeepMind
- Check Point Research
- Palo Alto Networks Unit 42
- Rapid7
- FortiGuard Labs
- Krebs on Security
- MIT CSAIL
- Nextgov
- Allen Institute for AI (Ai2)
- The Register
- Rest of World
- FedScoop
- Electronic Frontier Foundation