# PHP Ships Again While the Inventory Stays Half Red

Published: 2026-11-19
Canonical: https://agidreams.us/edition/php-ships-again-while-the-inventory-stays-half-red
Content-Complete: true

<!-- SECTION: 🐘 PHP Ships Again While the Inventory Stays Half Red -->

The PHP security team pre-announced another round of security releases for Thursday, 19 November, between 14:00 and 14:30 UTC, covering every supported branch as part of the regular four-week cycle. The notice is terse by design: no CVE identifiers, details deferred to release time, and one sentence that matters for advance triage, that "the issues rated CRITICAL require either non-default configuration or uncommon setups." The archive now lists PHP 8.5.13, 8.4.28, 8.3.37 and 8.2.36, each described only as "a security release." That is a different shape of problem from the out-of-cycle 21 October release, which fixed two CVSS 9.2 unauthenticated remote code execution bugs, CVE-2026-109618 and CVE-2026-109619, with the guidance that "there are no complete mitigations, updating is the only safe choice." One loose end is CVE-2026-115560, an "unlikely RCE vector" for which PHP promised a full advisory "with the next release." Whether today's releases close it is not stated in the archive text; the release notes would settle that. PHP 8.2 reaches end of life on 31 December (more: https://www.php.net/archive/2026.php#2026-11-18-1).

The better measure of where the October bugs stand is the Dutch NCSC's php2shell repository, which tracks vendor status product by product in SPDX files. As of 12 November, three weeks after the fix, it listed 972 products: 213 fixed, 212 still vulnerable, 92 under investigation, 240 "runtime-dependent" (the product inherits whatever PHP the administrator installed) and 214 not affected. Fixed and vulnerable are tied. PHP is "embedded in appliances, network devices, management interfaces and other vendor products where administrators may not be able to update PHP directly." The advice to keep `session.upload_progress.enabled=0` after patching stands until the promised advisory lands (more: https://github.com/NCSC-NL/php2shell).

The nginx HTTP/2 flaw CVE-2026-123007 now has public proof-of-concept code. Rapid7's Stephen Fewer published one demonstrating code execution against nginx in a local Docker container, with constraints stated up front: roughly 128 GB of network traffic and one specific processor-dependent build of a standard library routine. Two replies ask whether it works against a server carrying real traffic, and one reports that it did not. A public demonstration against a loaded target would settle it, and nobody should wait for one before patching (more: https://x.com/stephenfewer/status/2123081992799453221).

The JVM's AVX-512 Base64 decoder bug has an awkward paper trail. OpenJDK pull request 32781 describes a heap-corrupting write in the x86_64 vectorized decode path, reachable only with malformed input, and proposes a small early-return fix. The page then shows a bot warning that the PR had been "inactive for more than 8 weeks," cross-references from Adoptium crash reports, and a closed PR with an outsider asking, "just curious, why was this closed?" No answer appears. What the page establishes is that a memory-corruption description for this code path sat in a public tracker for months, filed as a crash bug (more: https://github.com/openjdk/jdk/pull/32781).

A LinkedIn data exercise asks whether public exploit code correlates with a CVE being exploited, or exploited faster. The first pass, stitched from many sources by the author's own admission, says: "Most CVEs with public code show no exploitation, and most exploited CVEs had no public code when attacked." A commenter who works takedowns adds that phishing and leaked credentials are efficient enough that exploits are "the last thing we put energy into." Public code is a weak predictor of attacks, and so is its absence (more: https://lnkd.in/p/gUAAavNH).

<!-- SECTION: 🧱 A Patch-Born KVM Escape and a TPM That Forgets -->

Terra Security disclosed CVE-2026-99312, a guest-to-host escape in KVM, the hypervisor built into the Linux kernel, affecting AMD hosts with nested virtualization enabled. By Terra's account it is demonstrated: from an ordinary DigitalOcean droplet, needing only root inside their own VM, they read the host's kernel version banner and then about 127 GB of host RAM, mapped read-only. The write half was shown only on Terra's own bare-metal AMD server, where a guest changed 32 bytes in a host kernel page. The bug class is a use-after-free in how KVM tracks a guest's memory map when a nested guest is reset. The provenance is the part worth dwelling on: the flaw "sat in the kernel for just under four years, and it was opened by a security fix," a 2022 patch that Terra says has now needed at least two follow-ups.

For operators the facts are short. The fix was first tagged in v7.4-rc2 on 9 November and is in stable 7.3.5 and long-term kernels 6.18.64 and 6.12.118. Hosts that cannot update can set the kvm_amd module's nested parameter to 0, which Terra says closes the path completely. Exposure is narrower than the headline suggests: many providers Terra surveyed were out of reach because they never enabled nested virtualization on AMD. It includes no statement from DigitalOcean and no report date, so the disclosure timeline is incomplete (more: https://www.terra.security/blog/kvm-guest-to-host-escape).

A different trust boundary takes a hit in an arXiv paper, an updated version of an ACM CCS '26 publication, on the hardware that keeps Windows Hello and BitLocker PINs from being guessed. Firmware TPMs from AMD, Intel and Microsoft store their lockout counters on SPI flash, and newer platforms rely on Replay Protected Monotonic Counters (RPMC), a JEDEC feature inside the flash chip, to stop an attacker restoring an old copy with a fresh guess budget. The authors call theirs the "first practical security analysis of RPMC in shipping x86 systems," and conclude the ecosystem "is not up to its task." Across nine systems they report weaknesses in a flash vendor's key protection, in AMD's handling of counter updates, and in AMD desktops with no RPMC flash at all.

The measured results make this more than a lab curiosity. On an AMD Copilot+ laptop, a 4-digit Hello PIN falls in a worst case of 40 hours instead of 69 days. On an Intel Framework Laptop 13, BitLocker's default 6-digit TPM+PIN falls in 20 days instead of 19 years. Vendor response is the weak spot: "none of them has committed to fixing the root-cause vulnerabilities." Winbond, first told in January 2025, is evaluating a possible silicon revision with no timeline; AMD acknowledged the report without an assessment; Microsoft closed the Hello case as moderate severity with no CVE, but will raise the recommended minimum PIN to 6 digits. The practical advice is longer PINs (8 digits for BitLocker pushes the worst case to six years). Stated limitations: an unoptimized setup, a small sample, and no ARM systems tested (more: https://arxiv.org/abs/2611.18126).

<!-- SECTION: 🐤 Two Million Canaries in the Inference Supply Chain -->

HD Moore, creator of Metasploit and now research chief at Dragos, published a five-month experiment that turns "who is holding your prompts" from a contract question into a measurement. From July through November his team sent roughly two million prompts to widely used AI services, each seeded with unique URLs, DNS names and real AWS keys tied to empty, fully logged accounts, then watched where the secrets surfaced. The post reports "security breaches, prompt theft, and widespread failures to protect customer data."

Three unnamed hosting providers were being actively robbed, the earliest since mid-July, and "none of the three providers knew about their breach until we notified them." In the largest case every request in certain windows was stolen, including requests rejected before a model ran, which points to copying at the provider's front door. In the other two, some stolen requests carried Zero Data Retention (ZDR) flags: "A privacy flag doesn't help when the attacker is already inside." One adversary appears to have held two providers at once, and on 1 November added ARTEX, the LLM-driven penetration testing framework, to its tooling.

The quieter findings may matter more because they need no breach. Sellers on an inference marketplace captured prompts by listing models at the lowest price and turning on request logging. An unnamed Chinese lab fetched secret URLs from more than half the prompts sent to its models, including roughly two-thirds of those marked no-data-collection or ZDR, and prompts sent to U.S. resellers of those models triggered the same scraper. One link from a ZDR-flagged prompt sent to a U.S.-based provider turned up 24 days later in a Feishu chat, opened from six devices in Beijing and Shanghai. All of this rests on Dragos's own telemetry; the providers are unnamed and only one has published a preliminary notice. Moore's recommendations are unglamorous: self-host if the budget allows, go directly to a first-tier provider, avoid reseller marketplaces, and plant a fake credential in agent context so that any use of it is an incident. His summary: "a privacy flag only binds the first hop" (more: https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens).

Against that backdrop, a launch post from adverserialAI reads as a test case. The lab, self-described as "the best AI lab in the US specialized in cyber capable models," announced CyberSEEK, a fully abliterated model (refusal behavior removed) served from Intel TDX confidential VMs with "True ZDR. no logs, no traces, no training on your data," at $0.50 per million input tokens and $2 output. Confidential computing is an architectural claim, stronger than a policy flag, but still an assertion in a post. The post does not mention attestation a customer can verify; that, or a canary run of the kind Dragos describes, would move it from asserted to corroborated (more: https://x.com/lordx64/status/2123124861899825480?s=20).

<!-- SECTION: 🕷️ ARTEX: Boring Techniques, Machine Cadence -->

ReversingLabs published the most complete defender's map yet of ARTEX, the open-source agentic offensive framework released in July by a developer known as Autumn-27 for authorized testing and pulled on 8 October after abuse reports. Forks and archives kept it alive. Researcher Robert Simmons counts 11 fork families across 3,096 actionable IP indicators and 184 operator binaries for Linux, macOS and Windows, all importing the same orchestration library. The central observation is deflationary: "The concepts implemented underneath are remarkably boring and bog standard." No zero-days, just known weaknesses chained without the pauses a human operator takes. "Agentic frameworks remove that grace period."

The detection advice follows from that. The best pivot was the management console's favicon, which led to 969 confirmed live consoles. On the wire, the tell is cadence: bursts of process and network events separated by model "thinking" pauses, randomized scanning that slips under legacy thresholds, DNS bursts with NXDOMAIN spikes, and machine-speed SSH logins followed instantly by commands. LLM gateways deserve their own alerts on login-failure spikes, and coding-agent credential directories should be kept off exposed paths. The caveats are candid: defenders "will rarely observe the ARTEX binary itself on victim endpoints," User-Agent strings can be overridden, and some listed addresses are shared VPN or cloud exits, so the lists suit retroactive hunting, not a permanent blacklist. Defending "does not require futuristic AI countermeasures" (more: https://www.reversinglabs.com/blog/threat-researchers-notebook-detection-strategies-for-artex-and-related-forks).

A third-party explainer site, built from the RuoJi6 fork's README and code, gives the product view: one Go binary with an embedded web console over PostgreSQL, a planner dispatching work to three default workers running ordinary tools, and support for any Anthropic- or OpenAI-compatible endpoint. The README says every tool call passes rules, an LLM judge, then a human; the same page notes that a rules-of-engagement scope check "was removed, and ThreatBook says the open edition has no built-in safety controls" (more: https://www.artex.sh/how-it-works). Its companion tracker makes a point worth keeping in view: victim counts "differ by outlet and date," so each figure is what one source said on one day (more: https://www.artex.sh/news).

OrcaRouter used the moment to launch Shield, which answers offensive agents with deception instead of blocking. The company says it pointed ARTEX at its own production and lost real credentials from exposed configuration files within five minutes; Shield now serves probes a synthetic world of fake endpoints and decoy keys while humans get the real site. A reply argues it works only against rigid probes and would last "a week." An independent test against an adapted operator would settle that (more: https://x.com/orcarouter/status/2123258076921053596).

At the small end, a hobbyist released CyberQwen-2B, a fine-tune for security triage and CTF help that runs offline at 12 to 15 tokens per second on an 8 GB CPU-only laptop. The privacy rationale lands well this week. The evidence does not: commenters ask whether it was run on known cyber evals, and one notes that models this size "just hallucinate mostly" (more: https://old.reddit.com/r/ollama/comments/1xw84q1/i_finetuned_a_2b_model_for_cybersecurity_ctf/).

<!-- SECTION: 💾 180 Billion Parameters, 32 GB, and an Honest Scorecard -->

A Hugging Face engineer set out to run Qwen3.8-Flash-Next, a 180B-parameter mixture-of-experts model (MoE, where a router activates only a few "expert" sub-networks per token), on a 32 GB MacBook through stock Transformers, and wrote it up with unusual candor. The benchmark to beat was Strata, the specialized engine that keeps hot experts in fast memory and spills the rest to RAM or disk. The diagnosis of existing options is crisp: llama.cpp is "hard to beat" when experts fit in RAM, otherwise "the page cache thrashes"; vLLM has no disk tier; and Transformers copies a layer's entire expert bank per token "to use 8 or 10 of them."

The fix is a single preallocated device tensor shared by all layers, with a per-layer table mapping expert IDs to slots so the pool "looks like a regular expert bank." With experts quantized to 3 bits (49 GB on disk, 12 GiB cached) and the n-gram embedding table left on disk, memory peaks at 25.4 GiB on a 6K prompt. Speed went from about 10 to about 14 tokens per second, roughly 6% behind Strata on short decode. Then the scorecard: "Transformers didn't beat Strata," the result is "still too slow for how I use coding agents," and "I have to close the browser, or macOS starts swapping." All of the code was AI-written (more: https://huggingface.co/blog/moe-expert-streaming).

The same split shows up in a community quant of Tencent's Hy4-flash, a 220B model with 155.5B in the network (512 routed experts, 16 active) and 63.5B in a hashed n-gram table that stays on SSD. Experts at about 2.3 bits bring in-memory weights to 49.1 GB, down from the 72 to 105 GB of Tencent's own builds, enough for a 64 GB machine. No quality numbers accompany the quant (more: https://old.reddit.com/r/LocalLLaMA/comments/1xzopsy/hy4flash_quants_are_down_from_70_to_49_gb_run_on/).

Thinking Machines released Inkling-2-Small, and the thread is mostly about the adjective. The post has no body; commenters cite 383 billion parameters in one place and 600B in another, so the model card is the authority. One commenter flags a license clause under which continued use constitutes acceptance of any updated acceptable-use policy, a real consideration for anyone building on it (more: https://old.reddit.com/r/LocalLLaMA/comments/1xuqplg/thinkingmachinesinkling2small_is_released/).

At the other end, a user reports SenseNova-Loop-9B, a looped transformer that reuses its layers, running at 300 to 500 tokens per second on a laptop RTX 5090 with NVFP4 weights, aggregated across eight parallel agents. Against Qwen3.8 27B on tool-eval-bench it scores 90 versus 98 on standard tests and 50 versus 100 on the hard injection category. The author allows that quantization may have degraded it. Half the injection resistance is a poor trade for speed in anything that touches untrusted input (more: https://old.reddit.com/r/LocalLLaMA/comments/1xz3f4w/running_sensenovaloop9b_nvfp4_looped_transformer/).

<!-- SECTION: 💡 Gemini 4 Neon and the Environment Built for Agents -->

Google introduced Gemini 4 Neon as the middle of its Gemini 4 family: "While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency." Pricing is $3 per million input tokens, $0.15 cached and $15 output, with a 1M-token context. The headline claims are first place on the Artificial Analysis Coding Agent Index (58.6), 46.2% on Cognition's FrontierCode 2.0, and 32% lower cost per task than Carbon. Every number is relayed by Google, including those from third-party benchmarks, and one result shows how much the harness matters: 12.8% on SWE-sweep with a minimal harness versus 39.0% inside Google's own Antigravity.

The safety figures deserve the same reading. Google reports a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark, but with real-time safeguards at the highest thinking level, which is not the default. It also reports shortcuts on 2.45% of tasks in its coding cheating evaluation. Independent reruns at default settings would confirm or refute these (more: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/).

Models are increasingly consumed through orchestration layers, and Orca pitches itself as an "ADE," an agent development environment: Claude Code, Codex, OpenCode, Grok or any agent CLI in parallel, each in its own git worktree, with an embedded browser, diffs and a mobile companion. It is open source. The design choice that matters is that "anything you can do in Orca, your agents can do too," via a CLI: create worktrees, start workers, drive desktop apps, schedule automations. The page does not describe a permission model for agents that can drive a desktop and schedule their own work, which is the first thing a security reviewer will ask (more: https://orca.dev).

<!-- SECTION: 🎬 Films From Code, Sound by Description, Speech in 4 MB -->

cinecode is an MIT-licensed skill kit that lets a coding agent deliver a finished MP4 with no video model involved. The agent writes the script, builds each frame in code, renders in headless Chrome and encodes with ffmpeg; "every frame is a pure function of time," so re-renders are identical. The discipline around it is the interesting part: documentary claims must be sourced, archive material is logged in a licence file, and guard hooks cap each helper sub-agent at 25 steps. Limits are stated: no end-to-end Windows test yet, and a 30 to 60 second film takes 100 to 250 agent steps (more: https://github.com/sarpstar/cinecode).

AudioGhost v1.2 wraps Meta's SAM-Audio in a local web app: type "background music" or "dog barking" and it isolates or removes that sound from audio or video. The small model fits in a bit over 4 GB of VRAM because the author stripped the official pipeline's helper models, and says what that costs: quality "a bit below the full pipeline" and results that vary run to run. A commenter asks whether a half-caught sound leaves a muffled ghost or a hole in the dialogue (more: https://old.reddit.com/r/LocalLLaMA/comments/1xz6xl7/audioghost_v12_remove_any_sound_from_audio_or/).

Kitten-TTS-2 pushes small speech models further down: Pico at about 7M parameters and 4 MB, Nano at 17M and 8.8 MB, both covering 8 languages, and Mini at 80M and 48 MB with 16, all Apache 2.0. The developers say Pico runs faster than real time on a single smartphone CPU thread. Early testers call the quality good in English and German, with one concrete defect: it ran sentences together, "almost completely (but not quite) ignoring commas and full stops" (more: https://old.reddit.com/r/LocalLLaMA/comments/1y0e59m/kittentts2_pico_nano_mini_sota_supertiny_apache/).
