You're reading the Monday, November 16, 2026 edition. Today's briefing →
Toronto
Monday, November 16, 2026No. 155

Digital Plumber

Plumbing the information age

AI-curated intelligence for people who run networks. Daily coverage of AIOps, network automation, agentic operations, AI infrastructure, security and the vendors shaping them.

Today's 3 things that matter

Picked by the AI editor
  1. Startups·Industry news

    Traversal raises $120M for autonomous AI site reliability engineers

    Traversal, which builds autonomous AI site reliability engineers to diagnose and resolve software outages, raised $120 million in Series B funding from Kleiner Perkins, Redpoint Ventures, Sequoia Capital, Battery Ventures, and Khosla Ventures at a $1 billion valuation.

    Why it matters Demonstrates market traction for autonomous incident response; offers practitioners an alternative to traditional runbook automation and manual triage during outages—reducing MTTR and operational toil.

  2. Routing·Primary source

    FRRouting bgpd flaw crashes all BGP sessions

    A flaw in FRRouting's bgpd component allows improper validation of the data field length in a BGP NOTIFICATION Cease/Hard Reset message, existing in FRRouting releases 8.3 through 10.5.1.

    Why it matters Affects operators running FRRouting as their BGP daemon; allows remote DoS via crafted NOTIFICATION messages from established BGP peers with Graceful Restart capability.

  3. DC networking·Industry news

    Lambda runs co-packaged optics across 10,000 Blackwell GPUs

    Lambda and NVIDIA announced the first large-scale commercial deployment of co-packaged optics (CPO) networking, connecting over 10,000 Blackwell Ultra GPUs at Lambda's Quebec data center.

    Why it matters CPO switches operating at 800 Gb/s achieve median optics temperature of 36°C versus 51°C for pluggables, reducing thermal density and enabling denser AI fabrics without data center infrastructure redesign.

Today's briefing

What happened, and why it matters

17 stories · 10 topics · Updated 3:20 PM ET

Ten DevOps tools that anchor a network automation pipeline

Top DevOps Training Institute Blog · Nov 9, 2026 · Industry news

Templates in Git render data into device configs; CI pipeline runs Batfish offline validation and pyATS checks before lab deployment and production rollout. Covers the complete NetDevOps workflow from source of truth through multi-stage testing.

Why it matters Application teams can open PRs against network repos instead of tickets, reducing wait time while network team retains approval authority.

This practitioner guide maps the full NetDevOps workflow: intent stored in a source of truth (devices, interfaces, IPs, VLANs), Jinja2 templates rendering data into configs in Git, pull requests triggering CI validation, offline Batfish analysis catching policy violations and routing errors, Containerlab lab deployment for real control-plane testing with pyATS validation, and production deployment with optional dry-run. The article positions network automation as part of broader DevOps: when application teams need VLAN, firewall, or load-balancer changes, they open PRs against the network repo instead of filing tickets. Network team retains approval but waiting time drops because code review, CI testing, and audit trails apply uniformly. Includes concrete .gitlab-ci.yml examples showing batfish_checks.py and pyats_checks.py pipeline stages, addressing the persistent challenge of integrating network changes into application-driven release cycles without manual handoffs.

Read the original at devopstraininginstitute.com ↗

OpenTelemetry graduates from CNCF with stable Kubernetes processor

XenonStack, Apica.io, DEV Community · Nov 12, 2026 · Primary source

OpenTelemetry reached CNCF Graduated status on May 11, 2026. As of November 2026, traces and logs are stable, continuous profiling entered public alpha on March 26, 2026, and the Kubernetes Attributes Processor reached v1.0.0 Stable status in September 2026.

Why it matters OpenTelemetry is now the de facto standard for telemetry collection across all major cloud providers and observability vendors (Datadog, Honeycomb, Grafana, New Relic). Network teams can standardize instrumentation across domains without vendor lock-in.

OpenTelemetry's graduation to Graduated status (the highest CNCF maturity level) alongside Kubernetes and Prometheus signals the completion of the industry-wide shift from vendor-proprietary instrumentation to vendor-neutral telemetry. The stable Kubernetes Attributes Processor (v1.0.0 as of September 2026) enables automatic collection of pod and node metadata without code changes, critical for correlating network events with container orchestration state. Continuous Profiling entered public alpha in March 2026 and is still not recommended for critical production workloads, but represents the fourth signal alongside metrics, logs, and traces. For network and infrastructure teams, this means a unified OTLP wire protocol for all telemetry, elimination of format translation overhead, and seamless correlation across network paths, container scheduling decisions, and application performance without tool-specific plugins. Vendor-sourced OTel distributions increased from 44% to 60% year-over-year, indicating platform adoption is accelerating beyond early adopters.

Read the original at xenonstack.com ↗

Lambda runs co-packaged optics across 10,000 Blackwell GPUs

CASRAI · Nov 16, 2026 · Industry news

Lambda and NVIDIA announced the first large-scale commercial deployment of co-packaged optics (CPO) networking, connecting over 10,000 Blackwell Ultra GPUs at Lambda's Quebec data center. The CPO fabric delivered 3.8% lower network latency at rack scale and 10.1% lower latency at pod scale compared to conventional pluggable optics.

Why it matters CPO switches operating at 800 Gb/s achieve median optics temperature of 36°C versus 51°C for pluggables, reducing thermal density and enabling denser AI fabrics without data center infrastructure redesign.

Lambda and NVIDIA announced a commercial CPO deployment at SC26 (November 16, 2026) connecting over 10,000 Blackwell Ultra GPUs in Quebec, Canada—the world's first verified commercial CPO deployment in an AI cloud environment. Co-packaged optics integrates optical engines directly onto the switch substrate beside the ASIC. Lambda's testing showed 3.8% lower rack-scale latency and 10.1% lower pod-scale latency versus pluggable optics with identical throughput. Each switch has 144 ports at 800 Gb/s with 18 removable external light sources, achieving median optics temperature of 36°C (versus 51°C for pluggables) and a PUE of 1.2 with water-free cooling. For operators: this is the first production verification that CPO deployments close the latency gap with pluggable optics while delivering the power and thermal benefits promised at smaller scales. CPO remains only ~0.5% of AI data center optical modules in 2026, rising toward 35% by 2030, so this deployment signals production readiness ahead of mass adoption and shifts the dial on feasibility for large-scale AI cluster fabric architectures.

Read the original at casrai.org ↗

Broadcom shows Tomahawk 6 and Thor Ultra at SC26

CASRAI · Nov 15, 2026 · Industry news

Broadcom demonstrated Tomahawk 6 and Thor Ultra 800G Ethernet silicon at SC26, highlighting compatibility with Ultra Ethernet Consortium (UEC) specifications for AI workloads. Thor Ultra eRoCE NICs achieve job completion times matching 800G InfiniBand quantum fabrics by placing packets directly into GPU memory.

Why it matters Open Ethernet silicon with UEC compliance and direct GPU memory placement eliminates packet ordering dependencies, allowing switches to spray packets across all links and closing the performance gap with InfiniBand at Ethernet scale and cost.

Broadcom showcased Tomahawk 6 switches (51.2 Tbps) and Thor Ultra 800G network interface cards at SC26, integrating emerging specifications from the Ultra Ethernet Consortium (UEC) and Open Compute Project's Scale-Up Ethernet Transport (SUE-T). Thor Ultra provides 800 Gbps per port with eRoCE support, direct GPU memory placement, and packet trimming capabilities, eliminating head-of-line blocking inherent in traditional RoCE. Technical analysis confirms that Tomahawk 6 combined with Thor Ultra achieves job completion times matching 800G InfiniBand Quantum fabrics. The silicon supports Credit-Based Flow Control and Link Layer Retry for intra-rack accelerator memory fabrics. For practitioners planning large-scale AI clusters: this demonstrates that open Ethernet fabrics have resolved the ordering and congestion issues that previously forced InfiniBand adoption, while maintaining multi-vendor sourcing flexibility and operational simplicity over proprietary solutions.

Read the original at casrai.org ↗

BGP hijack knocked out Cloudflare 1.1.1.1 in Canada

Cloudflare · Nov 14, 2026 · Primary source

A BGP route hijack of Cloudflare's 1.1.1.0/24 prefix disrupted the 1.1.1.1 public DNS resolver for users in Canada for approximately 1.1 hours. Cloudflare identified the offending network responsible for the route leak and contacted them directly to remediate the issue.

Why it matters Live route hijacks against major infrastructure remain common; this incident demonstrates both the attack surface and effective detection/mitigation responses practitioners should understand.

A BGP route hijack of Cloudflare's 1.1.1.0/24 prefix disrupted the 1.1.1.1 public DNS resolver for users in Canada for approximately 1.1 hours on November 13, 2026. Cloudflare reached out to the network where they saw an active hijack of 1.1.1.0/24, while 1.0.0.1 and public resolver IPv6 prefixes remained fully operational. The network with the offending route leak fixed the issue on their end, with Cloudflare monitoring the effect of the fix. The incident illustrates how more-specific route announcements exploit longest-prefix-match routing behavior to override legitimate announcements and that even large operators with mature BGP monitoring can experience customer impact during the detection-to-remediation window. For operators: this reinforces the value of RPKI validation and real-time BGP monitoring tools to alert on unauthorized announcements targeting critical infrastructure.

Read the original at isdown.app ↗

FRRouting bgpd flaw crashes all BGP sessions

NetDEF/FRRouting CNA · Nov 13, 2026 · Primary source

A flaw in FRRouting's bgpd component allows improper validation of the data field length in a BGP NOTIFICATION Cease/Hard Reset message, existing in FRRouting releases 8.3 through 10.5.1. This causes a heap buffer over-read and an assertion failure, leading to a crash of the bgpd daemon and denial of service by terminating all BGP sessions.

Why it matters Affects operators running FRRouting as their BGP daemon; allows remote DoS via crafted NOTIFICATION messages from established BGP peers with Graceful Restart capability.

In FRRouting (FRR) 8.3 through 10.5.1, bgp_notify_decapsulate_hard_reset() in bgpd/bgp_packet.c does not validate the length of the data field of a received BGP NOTIFICATION Cease/Hard Reset message (RFC 8538). The flaw occurs when processing a BGP NOTIFICATION Cease/Hard Reset message with a one-byte data field, causing a heap buffer over-read and an assertion failure, leading to a crash of the bgpd daemon and denial of service by terminating all BGP sessions. The vulnerability requires a remote attacker with an established BGP session and Graceful Restart capability negotiated with the Notification bit set. Upgrading to version 10.0.5, 10.1.5, 10.2.6, 10.3.4, 10.4.3, 10.5.2 or 10.6.0 eliminates this vulnerability. If upgrading is not feasible, disable the Graceful Restart BGP capability on both the local router and its peers using "no bgp graceful-restart notification" or "bgp graceful-restart-disable". Operators running FRRouting should prioritize patching to prevent bgpd crashes from established BGP peers.

Read the original at github.com ↗

Check Point Gaia OS root flaw exploited in the wild

CISA Known Exploited Vulnerabilities Catalog · Nov 13, 2026 · Primary source

CISA added CVE-2026-123039 to its Known Exploited Vulnerabilities catalog on November 13, 2026, confirming a Check Point Gaia OS flaw allowing local users to gain root access is under active exploitation. The CVSS 7.8 vulnerability affects the DynamicCLI service, escalating limited local access to full system control.

Why it matters Affects Check Point Force Firewalls, Management Servers, and Maestro Orchestrator; active exploitation confirmed against appliances in production. Requires immediate patching on all perimeter security infrastructure.

CVE-2026-123039 affects Check Point's DynamicCLI service in Gaia OS across releases R81–R82.20. The flaw resides in dynamic_server, a root-owned process that executes commands without proper authentication. Attackers have demonstrated privilege escalation chains via webshell installation in template directories, gaining appliance control. Check Point confirmed exploitation affecting approximately two dozen customers. Operators have installed PHP webshells and persistent backdoors (TSGA), confirming post-exploitation access. Organizations must apply patches immediately; CISA KEV deadline for federal agencies is November 16, 2026.

Read the original at cisa.gov ↗

Barracuda email gateways attacked via pre-auth root chain

BleepingComputer · Nov 10, 2026 · Industry news

Barracuda confirmed on November 9, 2026, active exploitation of two chained Email Security Gateway flaws (CVE-2026-116606, CVE-2026-111560) that allow unauthenticated attackers to execute OS commands as root. PoC exploit code publicly available since November 3.

Why it matters Any internet-facing ESG management interface is vulnerable to pre-authentication root compromise. Email teams must patch immediately—exploitation chain is public, confirmed active, and allows complete email infrastructure takeover.

CVE-2026-116606 (CVSS 9.3) is a pre-authentication authorization bypass in the ESG admin interface; Barracuda initially listed it as requiring admin access but corrected on November 5 to clarify no credentials are needed. CVE-2026-111560 (CVSS 8.6) is a command injection flaw executed as root. Together, they give unauthenticated remote code execution on any exposed ESG. Barracuda released patches on November 2 (BNSF-40425 for v9.4.017+, BNSF-40430 for earlier). watchTowr Labs published working PoC on GitHub around November 3. CISA added both to KEV catalog on November 10 with federal patching deadline of November 13. Exploitation confirmed affecting a small number of organizations as of November 9.

Read the original at bleepingcomputer.com ↗

OpenSSL discloses five low-severity QUIC and AES-GCM-SIV flaws

OpenSSL · Nov 10, 2026 · Primary source

OpenSSL published security advisory on November 10, 2026, documenting five low-severity vulnerabilities in QUIC (CVE-2026-95691, CVE-2026-88822), AES-GCM-SIV (CVE-2026-95692, CVE-2026-88823), and CMS (CVE-2026-107496). Fixes ready in GitHub; official releases scheduled December 1, 2026.

Why it matters QUIC use-after-free and AES-GCM-SIV issues affect all supported OpenSSL versions; organizations using HTTP/3, QUIC proxies, or edge infrastructure must plan library upgrades by December 1 release.

CVE-2026-95691 is a heap-use-after-free in the QUIC client triggered by a malicious QUIC server, potentially causing denial of service or information disclosure. CVE-2026-95692 is an out-of-bounds read and incorrect IV length handling in AES-GCM-SIV (affecting OpenSSL 4.1, 4.0, 3.5 only). CVE-2026-107496 is an infinite loop in CMS_encrypt() with an EC group order of 1, affecting 4.1, 4.0, 3.5, 3.0, 1.1.1, 1.0.2. CVE-2026-88822 is uninitialized stack bytes during QUIC peer detection (4.1, 4.0, 3.5). CVE-2026-88823 is AES-GCM-SIV AAD forgery after zero-length incremental update (4.1, 4.0, 3.5). All fixes are in the public GitHub repository; infrastructure teams running HTTP/3 endpoints, API gateways, or QUIC-accelerated services should prepare for patching by the December 1 release date.

Read the original at openssl-library.org ↗

Traversal raises $120M for autonomous AI site reliability engineers

Parsers Substack / Tech Funding Digest · Nov 10, 2026 · Industry news

Traversal, which builds autonomous AI site reliability engineers to diagnose and resolve software outages, raised $120 million in Series B funding from Kleiner Perkins, Redpoint Ventures, Sequoia Capital, Battery Ventures, and Khosla Ventures at a $1 billion valuation. The round reflects investor confidence in autonomous incident response and outage resolution.

Why it matters Demonstrates market traction for autonomous incident response; offers practitioners an alternative to traditional runbook automation and manual triage during outages—reducing MTTR and operational toil.

Traversal's $120 million Series B at a $1 billion valuation underscores strong investor conviction in autonomous incident response as a core infrastructure function. The company builds AI agents that autonomously diagnose and resolve software outages—moving beyond static runbooks or alert escalation to proactive problem-solving. This sits at the intersection of AIOps and DevOps: the platform learns from past incidents and organizational infrastructure patterns to reason about failures and execute remediation. For SRE teams managing complex, distributed systems, Traversal represents a new category of tooling that can operate during the most critical periods (outages) when human attention is most strained. The funding syndicate—including top-tier infrastructure VCs like Kleiner Perkins and Sequoia—signals this is not a niche play but a core capability expected to become standard in enterprise incident response. The $1 billion valuation reflects the high stakes: reducing mean time to resolution (MTTR) translates directly to revenue protection and customer satisfaction for enterprises.

Read the original at parsers.substack.com ↗

Q/C Technologies promotes Yossi Ehrlichman to CTO

Globe Newswire · Nov 16, 2026 · Vendor release

Q/C Technologies appointed Yossi Ehrlichman, Ph.D., as CTO in September 2026, promoted from his role as Founding Manager of PIC Development. Ehrlichman leads the company's technology roadmap for its optical processing unit (OPU) designed to accelerate AI computations using light.

Why it matters Optical processing architecture represents an emerging alternative to electrical AI infrastructure for reducing energy constraints in GPU clusters and large-scale AI deployments.

Q/C Technologies, which is developing an optical processing unit (OPU) targeting AI infrastructure energy constraints, formalized Yossi Ehrlichman's appointment as CTO in September 2026, promoted from his previous position as Founding Manager of PIC Development. Ehrlichman holds expertise in integrated photonics, electro-optic devices, and architectures for optical computing—critical for photonic AI acceleration. He is a named inventor on patents spanning electro-optic modulation, optical signal generation, and photonic manipulation. The appointment comes as Q/C Technologies has assembled a photonics engineering team and established laboratory facilities in San Diego. The company announced over $10 million in cash and investments to fund OPU development. For infrastructure operators and NetDevOps teams, optical processing represents a potential paradigm shift in AI cluster networking, addressing power efficiency limitations of electrical systems.

Read the original at globenewswire.com ↗

Active MCP servers pass 118,000 as licensing gaps persist

DEV Community · Nov 14, 2026 · Industry news

Catalog snapshot on 14 November 2026 shows 118,434 active MCP servers, with 62,952 listed in the official MCP Registry. 15,380 servers and 7,500 skills had recent activity reports; 13.1% of skills carry no license.

Why it matters The scale of available MCP infrastructure—118k servers—means ops teams must establish governance policies for which servers are trusted before deploying agents to interact with network, cloud, and observability systems.

The HostedMCP database on 14 November 2026 counted 118,434 active MCP servers, with 62,952 published through the official MCP Registry. More than half of all servers now come through the official registry since its 2025 launch, though 13,357 servers flagged from bulk publishers indicate automated generation. Among the ecosystem, 15,380 servers and 7,500 skills had recent activity reports; 6,853 of 52,147 skills (13.1%) carry no license, creating compliance uncertainty. Governance remains inconsistent across the ecosystem—critical for ops teams integrating MCP servers into agent-based workflows. The rapid growth (from near-zero in 2025 to 118k in November 2026) means infrastructure teams must implement vetting processes, capability inventories, and runtime access controls before connecting agents to production systems.

Read the original at dev.to ↗

Microsoft Defender now discovers local AI agents and MCP links

Microsoft Tech Community · Nov 16, 2026 · Primary source

Microsoft Defender now discovers more than 50 supported local AI agents and their configured Model Context Protocol connections. Generally available for Windows and macOS, it brings agent, device, account, and MCP context together by observing activity.

Why it matters Network and security ops teams can now use Microsoft Defender to discover which agents and MCP servers are running locally—essential visibility before scaling agents in production, especially for SOCs managing agent-to-infrastructure connections.

Microsoft Defender's agent and MCP discovery feature is now generally available for supported local agents on Windows and macOS. The discovery brings agent, device, account, and MCP context together by observing activity patterns. This is a significant operational milestone: agents are no longer invisible to security monitoring. Network ops teams can inventory which agents and tool servers are connected to their infrastructure, track MCP configurations, and enforce governance policies. Coverage currently extends to 50+ supported agents; not every agent or connection will be visible yet, but this provides a foundation for agent-aware security operations. For infrastructure teams running agents in production, this enables baseline visibility and audit trails of agent-to-system interactions.

Read the original at techcommunity.microsoft.com ↗

Agentic pentest framework ARTEX found on 1,465 hosts

Cloud Security Alliance · Nov 16, 2026 · Industry news

Cloud Security Alliance documented ARTEX agentic pentest frameworks now deployed at 1,465 hosts (as of early November 2026), with 10.6% also hosting Cobalt Strike, Vshell, and Adaptix C2. ARTEX automates vulnerability discovery and exploitation using multi-agent frameworks with no human approval in the loop.

Why it matters Network defenders must shift from signature-based detection to behavioral telemetry: ARTEX and similar frameworks turn multi-step attacks into machine-speed intrusions requiring detection of timing patterns, chained requests, and rapid bursts across systems.

Cloud Security Alliance published analysis of ARTEX agentic pentest frameworks found on exposed staging servers. Censys counted 1,465 hosts exposing ARTEX consoles as of early November 2026, with 10.6% also hosting other threat tooling (Vshell, Cobalt Strike, Adaptix C2). ARTEX organizes planner/worker agents that continuously diff patches, draft exploits, and execute them without human approval. The framework leaves identifiable request patterns in logs: specific tool chains, source-map reads, rapid password-spray bursts, and chained API calls. For network operations and incident response teams, this means traditional signature-based blocking and manual triage are insufficient. Detection requires behavioral correlation across multiple systems, understanding how agents sequence tools over time, and recognizing timing anomalies that differ from human operator patterns. Organizations must implement cross-system observability and behavioral baselines for agent-driven attack detection.

Read the original at labs.cloudsecurityalliance.org ↗

Nvidia CEO exits as Anthropic sets IPO date

AI Weekly · Nov 16, 2026 · Industry news

In the first half of November 2026, Nvidia's CEO Jensen Huang resigned, Anthropic set its IPO for November 19, Anthropic and Google shipped new frontier models. Claude Fable 5.2 (November 9) is built for complex coding and research with 30% faster output at $8/$40 per million tokens; Gemini 4 Carbon (November 10) is Google's strongest model at $2/$10 introductory pricing.

Why it matters Fable 5.2 scored 49.1% on FrontierCode 2.0 at half the cost per task of Opus 5.5, directly affecting model selection economics and infrastructure budgeting for teams building agentic coding workflows.

Anthropic set IPO pricing for November 18 at $70–$78 per share, trading as ANT on November 19. Claude Fable 5.2 (released November 9) achieves 49.1% on FrontierCode 2.0 (highest of any Anthropic-tested model), matching Opus 5.5's best effort score (45.3% vs 45.5%) at roughly half the cost per task. Output is up to 30% faster than Fable 5.1; pricing is $8 input and $40 output per million tokens. In testing, Fable 5.2 coordinated over 1,000 subagents for eight days to build a bootable operating system. Google released Gemini 4 Carbon on November 10 as its strongest model after a two-week defensive cybersecurity preview. Carbon features a unified 1M-token context window shared between input and output (versus separate windows in earlier Gemini models), supports text, image, audio, and video input, and includes tunable thinking levels (low, medium, high). Introductory pricing through November 30 is $2 input and $10 output per million tokens, rising to $4/$20 from December 1, with 95% cache discount. Carbon is based on Gemini 4 Argon and uses more tokens for reasoning and self-verification than its predecessor.

Read the original at joinaiweek.com ↗

Companies borrowed $680 billion for data centers this year

Axios · Nov 16, 2026 · Industry news

Companies borrowed more than $680 billion to finance data centers and other advanced computing infrastructure in the first 10 months of 2026 — roughly five times the amount raised in all of 2023. 87% of Fed survey respondents cited AI as a risk to financial stability, up from 50% six months earlier.

Why it matters If AI investments fail to generate expected returns, losses could spread beyond technology companies to the banks and investors financing them. Banks have nearly $1 trillion in financial exposure to AI infrastructure.

Federal Reserve Governor Michael Barr on November 16 laid out perils AI could bring to trading, cybersecurity and consumer protection, with 87% of Fed contacts surveyed citing AI as a salient risk to the financial system. Barr pointed to cyber risks from AI-enabled cyberattackers and rogue AI agents exploiting vulnerabilities in financial systems. AI developers struggle with ensuring their models are well controlled, numerically precise, explainable and replicable. The Fed warned against risks from AI use in financial markets including collusion, market manipulation, or trading strategies causing significant volatility or systemic risk. In his Milan speech, Barr said autonomous agents might detect yield or stability signals and move retail deposits in minutes, overwhelming traditional liquidity controls. This represents the clearest public readout yet connecting AI capex, 1990s telecom parallels, and fire-sale risk.

Read the original at axios.com ↗

OpenAI embeds four outside evaluators with code access

CNBC · Nov 16, 2026 · Industry news

OpenAI signed agreements with four outside assessors—Apollo Research, METR, Transluce and Goodfire—for the first phase of embedded safety assessments covering model alignment, safeguards and monitoring, giving outside experts access to OpenAI's code and training data.

Why it matters Independent evaluators have taken on outsized importance following a series of high-profile safety and security incidents this year. OpenAI's move aligns with White House Accord on Super Intelligence commitments to rely on independent assessments of safety controls.

Apollo Research will place six evaluators inside OpenAI for four months to investigate whether AI models undermine their own or successor models' safeguards. METR will conduct a broader assessment of OpenAI's monitoring systems, with a preliminary agreement pending scope agreement. Transluce will examine the quality and coverage of OpenAI's post-training alignment evaluations, with plans for a follow-on assessment of eight weeks in early 2027. Assessors needing sensitive confidential information will work at OpenAI offices on OpenAI devices. OpenAI controls staff access and can redact some evidence but cannot hide assessment verdicts or withhold material because it is unfavorable. Disagreements escalate first to OpenAI's Safety Advisory Group and executives, then to the Safety and Security Committee of the OpenAI Foundation's board, which provides independent oversight.

Read the original at cnbc.com ↗
Nothing in today's briefing matches that.

Listening

Podcasts and talks
  • Elisaindustriq Blog (Polystar)

    TM Forum Autonomous Network Levels: Progress Toward Level 4

    Explores how telecom operators progress toward TM Forum Level 4 autonomous networks as traditional operational models based on manual processes become difficult to scale in complex, distributed, software-driven environments.

Vendor Radar

Last 7 days · arrows compare with the 7 before

What changed this week

Last 7 days vs the 7 before

Biggest moves

Trending topics

agent · observability · MCP · automation · Agentic AI · OpenTelemetry · Kubernetes · inference · RAG · AIOps · LLM · eBPF

Get the daily briefing

Today's 3 things that matter and every story with why it matters, in your inbox each morning. Free, and you can unsubscribe at any time. Prefer a reader? Follow the RSS feed.