/QAble Weekly/Vol. 019 · 30 Oct 2026

● This week’s signal » AI is shrinking the gap between a published fix and a working exploit. The gap that remains is how long it takes you to prove an upgrade is safe to ship.

Weekly

Signal Over Noise

‹ PrevNext ›

For Engineering
Leaders

5-minute read

‹ PrevNext ›

Friday, 30 October 2026 · Vol. 019

In Brief

CISA confirms PHP2Shell is being exploitedOct 26
Honeypots record the first confirmed PHP2Shell shellsOct 24
Microsoft removes trust in seven eMudhra certificatesOct 26
Microsoft offers voluntary buyouts to about 60,000 employeesOct 26
GitHub makes Copilot assisted permissions generally availableOct 26
Anthropic files its S-1 under the proposed ticker ANTOct 26
Apple launches its first touchscreen MacBook ProOct 28
CISA confirms PHP2Shell is being exploitedOct 26
Honeypots record the first confirmed PHP2Shell shellsOct 24
Microsoft removes trust in seven eMudhra certificatesOct 26
Microsoft offers voluntary buyouts to about 60,000 employeesOct 26
GitHub makes Copilot assisted permissions generally availableOct 26
Anthropic files its S-1 under the proposed ticker ANTOct 26
Apple launches its first touchscreen MacBook ProOct 28

Story of the Week

PHP was patched on Wednesday. By Saturday, attackers had working shells.

On 21 October, PHP released versions 8.2.35, 8.3.36, 8.4.27 and 8.5.12 to fix two critical flaws, CVE-2026-109618 and CVE-2026-109619, that allow remote code execution without authentication. Wiz Research, which disclosed them, says nearly every PHP version back to 5.4 is affected by at least one. It also says how it found them: it used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol throughout, and says Gemini 4 Argon made “an invaluable contribution to finding the vulnerability, exploit development, attack detection, and validation of the patch.” Its summary was blunt: “Frontier AI models substantially reduced that barrier, turning a longstanding flaw into an exploit.” On 22 October, the UK government said 32 of its services were unavailable and four partly unavailable, largely as a precaution while updates were applied. On 23 October, CISA said it had not observed successful exploitation. On 24 October, honeypots run by Perplexity’s Adel Karimi recorded 35 shells from two IP addresses. Expel’s Marcus Hutchins called it “by far the highest quality confirmation of in-the-wild PHP2Shell exploitation I’ve seen so far,” although Karimi cautioned that his deliberately favourable single-worker setup “is not a measurement of exploit reliability on other worker counts, PHP builds, or applications.” On 26 October, CISA added both flaws to its Known Exploited Vulnerabilities catalogue, with a federal deadline of 29 October. Censys counts more than 32.51 million PHP services still running an affected version.

Why it matters: Three days from patch to confirmed shells is now a planning assumption, not an outlier. Upgrading a language runtime touches every application on it, so the speed of a safe upgrade is set by the speed and trustworthiness of your regression suite. Official status lags reality: “not observed” on Friday became “known exploited” on Monday.

CISA added both PHP2Shell vulnerabilities to its Known Exploited Vulnerabilities catalogue on 26 October, giving US federal agencies until 29 October to patch. · CISA, via The Hacker News

WeeklySection 01 · This Week’s Launches

Product Launches

GitHub lets a model approve your agent’s commands, and says plainly that it can be wrong

What: GitHub now lets a language model approve an agent’s commands, and its own documentation says the model can approve dangerous actions.

On 26 October, GitHub made assisted permissions generally available in Copilot CLI, the Copilot app and Visual Studio Code. When an agent asks to run a command, access a file path or open a URL, existing rules and earlier human approvals are checked first. Anything still unresolved and eligible goes to a language model, which judges whether the action is safe and within the authority you gave for the task. If it is confident, the action runs without a prompt; if not, the user is asked. Enterprise deny rules always apply, and administrators can block the “Allow All” mode while keeping the assisted one. The part worth reading twice is in GitHub’s own documentation: “The model can make mistakes and approve dangerous or unintended actions.” And: “This feature is not a substitute for sandboxing or access restrictions.” That candour is the right design. A model is a filter that reduces prompts, not a boundary. On 23 October, Dynatrace detailed a Vulnerability Risk Exposure dashboard that separates vulnerabilities from the processes they affect and flags monitored processes that are not being scanned, with a line every security dashboard should carry: “A low vulnerability count can reflect limited scan coverage rather than limited exposure.”

Why it matters: A model approving another model’s actions reduces friction, but the real boundary is still the sandbox and the permissions underneath it. Vendors stating the failure mode in their own documentation is a good sign. Read it before you enable the feature. A low count on any security or quality dashboard means nothing until you know what the scanner could not see.

Launch Log

GitHub Copilot assisted permissions
Generally available on 26 October in Copilot CLI, the Copilot app and VS Code: a model approves agent actions it judges safe and asks the user about the rest.
Dynatrace Vulnerability Risk Exposure
A dashboard that separates vulnerabilities from the processes they affect and flags monitored processes that are not being scanned.
GitHub Agentic Workflows
Made generally available on 28 October, two days after assisted permissions.

WeeklySection 02 · Frameworks & Failures

Frameworks

Your patch window is now set by your test suite

PHP2Shell turned a familiar question into an urgent one: if a critical runtime fix ships this morning, how quickly can you put it into production and know nothing broke? Three days separated the patch from confirmed shells this week, and frontier models are making that gap shorter, not longer. Four capabilities decide where you stand. First, know what is running where. You cannot patch a version you have not inventoried, and Censys still found more than 32 million exposed PHP services. Second, be able to upgrade a runtime in a day. That means a regression suite fast enough to run on demand and trusted enough that a green result is a decision, not a debate. Third, measure coverage before you trust a count. As Dynatrace put it this week, a low vulnerability count can reflect limited scan coverage rather than limited exposure, and the same is true of a low defect count. Fourth, test the assumptions you inherit. eMudhra’s keys could be guessed in 256 tries and were trusted for roughly 18 months because nobody downstream checked. None of these is new advice. What changed this week is the price of not having them.

Why it matters: The organisation that can test an upgrade in hours patches in hours; the one that needs a week of manual regression is exposed for a week. Inventory, coverage and regression speed are quality engineering capabilities, and they now decide security outcomes. Pick one runtime in your estate and time how long a safe upgrade would really take. That number is your exposure window.

Failures & Data

A certificate authority’s keys could be guessed in 256 tries, and were trusted for 18 months

Eight intermediate certificate authority keys at the Indian CA eMudhra, six used for TLS and two for code signing, were generated in April 2025. According to BleepingComputer, they “had values smaller than 256, allowing anyone to recover them by testing just 256 possible numbers.” Microsoft’s security response team says the recoverable keys could have been used to issue certificates that browsers and operating systems would treat as legitimate. Microsoft blocked the affected intermediates on 20 October, then on 26 October shipped an out-of-band Windows update removing seven certificates from its trust list, six eMudhra G3 roots and an AAA Certificate Services cross-sign, and did the same in Edge on 27 October. eMudhra revoked the eight intermediates, and Mozilla, Google and Apple began revoking them too. Microsoft was careful about two things. “The issue arose in eMudhra’s certificate hierarchy; it was not a compromise of Microsoft systems.” And: “Removing a root from a trust list is not the same as revoking the known impacted intermediate certificate authorities.” Strip away the cryptography and this is a testing failure. Whether a key is unguessable is one of the easiest properties in security to check, and for roughly 18 months these keys sat inside chains of trust that billions of devices rely on.

Failures & Incidents

PHP2Shell confirmed exploited in the wild (Oct 24 to 26)
Honeypots recorded 35 shells on 24 October, and CISA added both flaws to its Known Exploited Vulnerabilities catalogue on 26 October.
SecurityWeek and CISA
PHP2Shell exploit attempts began before the patch
Attempts started before fixes were released, after some details leaked, and millions have been recorded since.
SecurityWeek, Oct 26
eMudhra CA keys recoverable in 256 guesses (Oct 26)
Eight intermediate keys generated in April 2025 were blocked and revoked, and Microsoft shipped an out-of-band update removing seven certificates from Windows trust.
Microsoft MSRC and BleepingComputer

The Wider Beat

Microsoft offers buyouts to about 60,000 employees, roughly a quarter of its workforce

On 26 October, Microsoft’s chief people officer Amy Coleman sent employees a memo offering voluntary buyouts to about 60,000 people, roughly a quarter of the company, across the United States, the United Kingdom, Canada and India. According to KUOW, the offer covers selected staff at level 70 and below, the ceiling below vice president, with eligibility set by the head of each organisation. The sales organisation, some corporate services such as marketing, and some engineering teams are not taking part. Terms are comparable to Microsoft’s standard severance. A few months ago the company offered early retirement, on more generous terms, to long-serving employees at level 67 and below, and about 30% of those eligible took it; it expects lower take-up this time. Coverage was quick to link the move to the company’s spending on AI infrastructure, but a Microsoft spokesperson said AI is not driving the buyouts. For engineering leaders, the practical question is not why but who. Voluntary programmes are taken up by the people who believe they have good options elsewhere, and in long-lived systems those are often the people who know why the tests are shaped the way they are. When a large share of an organisation can leave at once, the knowledge that lives only in people’s heads becomes a delivery risk.

WeeklySection 03 · Editor’s Note

By the Numbers · The AI quality gap, quantified

QAble Weekly analysis · sources per figure

32.51M

PHP services still running a version affected by PHP2Shell, after fixes shipped on 21 October

Source: Censys, via The Hacker News, Oct 27

3 days

from PHP’s patch release to the first confirmed remote code execution, recorded on honeypots

Source: PHP and SecurityWeek, Oct 21 to 24

256

guesses needed to recover each of eight eMudhra certificate authority keys, generated in April 2025

Source: BleepingComputer, Oct 27

~60,000

Microsoft employees offered voluntary buyouts, roughly a quarter of its workforce

Source: KUOW, Oct 27

Editor’s Note

Viral Patel, Co-Founder of QAble — Viral PatelCo-Founder, QAble

“PHP shipped a fix on Wednesday. By Saturday, there were working shells on honeypots. On Monday, it was officially being exploited. How long would your upgrade have taken?”

How long would it take you to ship a runtime upgrade to production, with evidence that nothing broke?

Most of us have a mental model of patching that goes something like this. A critical fix lands, we assess it, we schedule it, we test it, and we roll it out over a week or two. That model assumed attackers needed weeks to turn a disclosed flaw into something that worked.

This week that assumption failed in public. The researchers who found PHP2Shell were explicit that frontier AI models helped them find the bugs, write the exploit and validate the fix. Their own summary was that AI turned “a longstanding flaw into an exploit.” Three days after the patch, a researcher’s honeypots recorded working shells. Two days after that, CISA listed the flaws as actively exploited.

I do not think the lesson is to panic about AI-assisted attackers. I think the lesson is that patch speed has quietly become a quality engineering problem. Upgrading a runtime like PHP touches every application on it. Whether you can do that in a day depends almost entirely on whether you have a regression suite that is fast, broad and trusted enough that a green run is a decision rather than the start of a discussion.

The other stories this week point the same way. eMudhra’s certificate keys could be guessed in 256 tries and sat in chains of trust for about 18 months, because a basic check was never run. Dynatrace reminded everyone that a low vulnerability count can simply mean the scanner could not see. And GitHub, to its credit, wrote down that the model approving your agent’s commands can approve dangerous ones.

All of these come back to the same unglamorous capabilities: knowing what you run, measuring what your checks actually cover, and being able to change something important quickly without breaking it.

So this week, pick one runtime in your estate and ask your team a simple question. Not whether you are vulnerable, but how long a safe upgrade would really take. That number is your exposure window, whatever the advisory says.

WeeklySection 04 · Briefing

Funding & M&A

Bedrock Robotics $600M · Series C
General Compute $100M · Series A

Research

PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet
Wiz’s disclosure is unusually explicit about the role of frontier models in finding the flaws, developing the exploit and validating the patch, and about what that means for how quickly attacks can scale.
Microsoft’s response to compromised eMudhra certificate authority keys
A clear account of the difference between removing trust and revoking certificates, and of what teams with machines that miss automatic updates still need to do.

Quote of the Week

“Frontier AI models substantially reduced that barrier, turning a longstanding flaw into an exploit.”

Wiz Research, on PHP2Shell, 21 October

Market Signals

01The interval between a published fix and confirmed exploitation fell to days this week, and the researchers involved credit frontier models with lowering the barrier.
02Official exploitation status is a lagging signal: CISA moved from “not observed” to “known exploited” in three days.
03Vendors are putting failure modes in writing, as GitHub did for assisted permissions and Dynatrace did for scan coverage.
04Inherited trust is under strain, from certificate authorities to runtimes, and downstream teams rarely test what they inherit.
05Large employers are reshaping their workforces at unusual scale, which turns undocumented engineering knowledge into a delivery risk.

Community & Debate

What does “not observed” mean?

CISA said on 23 October that it had not seen successful PHP2Shell exploitation, and confirmed it three days later. Teams relying on that first status to defer patching found out how quickly it can change.

CISA, Oct 23 and 26

How much evidence is enough?

Marcus Hutchins called the honeypot shells “by far the highest quality confirmation” he had seen, while the researcher who ran them stressed a deliberately favourable setup.

Marcus Hutchins, Expel, and Adel Karimi

Should a model approve an agent’s actions?

GitHub’s own documentation says its assisted permissions model “can make mistakes and approve dangerous or unintended actions”, which leaves teams deciding where the real boundary sits.

GitHub Docs

Elsewhere in Tech

Apple launches its first touchscreen Mac

An OLED MacBook Pro with touch support, on M5 Pro or M5 Max chips, at a $400 premium over the mini-LED models, shipping on 6 November alongside an M6 14-inch MacBook Pro and iMac.

MacRumors, Oct 28

Anthropic files to go public

Its S-1 was filed on 26 October under the proposed ticker ANT, with Goldman Sachs and Morgan Stanley leading. No price range or share count has been set.

IPOScoop, Oct 26

Guidewire to buy hyperexponential for $1.7B in cash

The insurance software maker is buying the London-based AI underwriting and pricing firm, with up to $400M more tied to recurring revenue targets.

Guidewire, Oct 27

Hackers post data after GMO refuses a ransom

Attackers behind unauthorised trades on GMO’s rented trading servers, which MetaQuotes says were reached through saved logins, published alleged customer data after GMO refused a demand of about ¥8 billion. GMO says up to 26,739 servers may have been exposed.

Finance Magnates, Oct 27

Discord will put public server messages in search results

Web pages for Discoverable servers, communities of 1,000 or more members, start on 16 November, and messages from eligible public channels may appear in search engines from mid-December, with opt-outs for admins and users.

Dexerto, Oct 27

Sources: 01 PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet (Wiz Research, Oct 21) | 02 Security advisory: critical remote code execution vulnerabilities, update now (PHP, Oct 21) | 03 Thirty-two UK government services offline amid PHP cybersecurity issue (PA, via AJ Bell, Oct 22) | 04 PHP2Shell exploitation moves from probes to working shells (SecurityWeek, Oct 26) | 05 CISA confirms active exploitation of two PHP2Shell flaws enabling remote code execution (The Hacker News, Oct 27) | 06 GitHub Copilot assisted permissions generally available (GitHub Changelog, Oct 26) | 07 Allowing safe commands (GitHub Docs) | 08 GitHub Agentic Workflows generally available (GitHub Changelog, Oct 28) | 09 SmartBear, Harness and Cognition: the latest vendor news (QA Financial, Oct 27) | 10 Microsoft’s response to compromised eMudhra certificate authority keys (Microsoft MSRC, Oct 26) | 11 Microsoft releases Windows OOB updates to distrust eMudhra certificates (BleepingComputer, Oct 27) | 12 Microsoft offers voluntary buyouts to a quarter of its workforce (KUOW, via OPB, Oct 27) | 13 Here’s everything Apple announced this week (MacRumors, Oct 28) | 14 The IPO Buzz: Anthropic (ANT proposed) files its S-1 (IPOScoop, Oct 26) | 15 Guidewire signs definitive agreement to acquire hyperexponential (Guidewire, Oct 27) | 16 Hackers publish alleged GMO customer data as broker FXTF shuts out overseas logins (Finance Magnates, Oct 27) | 17 Discord will soon let anyone find public server messages through Google (Dexerto, Oct 27) | 18 This startup’s self-driving excavators help dig up $600M in new funding (TechCrunch, Oct 26) | 19 General Compute raises $100M to finance inference beyond Nvidia (Runtime Wire, Oct 27)