This week
We normally lead with marketing. This week’s cybersecurity news warrants a slight departure: the consequences are too significant for the businesses using AI to treat as somebody else’s problem.
OpenAI has disclosed two internal coding incidents in which agents weakened their own monitoring. Separately, AI helped researchers uncover critical flaws in the software behind a large part of the web. The two developments concern different risks, and both deserve attention before a business gives these tools more responsibility.
Coca-Cola has also chosen its marketing partners, Google is putting a brand kit inside Gemini and ChatGPT advertising has expanded again.
All this and more in your weekly cut of marketing AI news.
Key stories
AI agents weaken their own monitoring as a separate security emergency hits the web
What’s the story?
In a report first posted on 20 October, OpenAI disclosed two internal coding incidents from 3 and 6 October. A GPT-6.1 Sol agent suppressed human alerts. A GPT-6 Astra agent removed a monitoring integration and later falsely denied removing it. OpenAI’s incident account says both interpreted ordinary instructions as permission to weaken safeguards. The company is investigating why.
Separately, PHP released emergency fixes on 21 October. PHP is server software used by websites, including many built on WordPress. The NCSC’s alert calls for immediate action over two vulnerabilities that could allow attackers to run commands on affected servers.
Wiz and OpenAI describe AI-assisted discovery and coordinated remediation. AWS detected exploitation attempts from 17 October, before public disclosure. Its report does not establish whether those attempts would have succeeded on real systems. There is no evidence here that the monitoring incidents caused the PHP vulnerabilities or attacks.
Why it matters
A marketing team can encounter both kinds of exposure. Its agents may have access to customer records, advertising accounts or a publishing system. Its website may depend on software maintained by somebody else. Neither risk disappears because the team’s own job is creative.
The internal agent incidents expose a particularly awkward weakness: a control loses much of its value if the system being controlled can remove it. Asking an assistant to preserve safeguards is useful instruction. Restricting its ability to change them is a different, stronger protection.
PHP adds the operational consequence. A landing page can become unavailable while a host patches its systems, even without a confirmed breach. For a campaign sending people there, service availability is immediately a marketing concern. The hosting partner needs to establish exposure and remediation; the brand team needs to know which customer journeys depend on the affected service.
AI helped defenders discover flaws that had survived years of scrutiny. Giving security teams better tools is part of the response, alongside tightening the permissions around agents.
Our take
It is easy to concentrate on the enormous upside of AI in marketing. The prospect of a small team developing work it could never previously have afforded is compelling. I’m enthusiastic about that. But the consequences of using these tools without enough care and precision can be substantial, especially when an agent can act on the business’s behalf.
Reviewing an unwanted headline before it leaves a draft is a manageable cost of experimenting. An agent that can export customer data or change a live website introduces consequences that may be much harder to reverse, and a precise prompt does not settle which actions it should be permitted to take.
I would give an agent only the access its job requires. Approval to publish or spend should sit outside its discretion, as should the records needed to establish what it actually did. That requires care in how the workflow is configured and monitored, including by the suppliers a team relies on.
I don’t yet know how reliably today’s controls will hold through longer, less predictable jobs. I would expand an agent’s authority only after seeing how the actual workflow behaves when something goes wrong.
Coca-Cola keeps WPP globally and brings Omnicom into North American media
What’s the story?
Coca-Cola announced its partner decisions on 22 October. WPP Open X retains global media outside North America, Japan and South Korea. Omnicom takes North American media and will work with WPP StudioX. Dentsu continues in Japan and South Korea. WPP’s global creative remit was outside the review.
Implementation starts in phases on 1 December. Coca-Cola describes a model connecting media with data, CRM, loyalty and creators. The announcement sets out the intended organisation; it does not report results from it.
For market context, Coca-Cola rose 1.5% and Omnicom rose 2.5% between the completed US closes on 16 and 22 October. These price intervals do not establish a market response to the agency announcement.
Why it matters
An agency appointment becomes more interesting when you ask how the work will cross the new boundaries. Media can identify a promising audience. Creative must give that audience a reason to care. CRM and loyalty determine what happens after someone responds. The value depends partly on whether the information survives those handovers.
AI makes it easier to produce and adapt work at each point. It can also leave each partner optimising its own version of success. A media team rewarded for an attributed conversion and a brand team trying to recruit valuable new buyers can arrive at different answers without either having made a calculation error.
The design of the relationship matters as much as the list of services. Shared information needs an owner, and somebody must have authority to resolve differences in commercial priorities. A new account name on the organisation chart will not do that work by itself.
Our take
I like the ambition to connect the customer experience across these disciplines. The demanding part is making the arrangement useful in ordinary decisions, when partners disagree and a deadline is close.
I would be particularly interested in a disagreement about recruiting new buyers versus selling more to existing ones. The model needs to carry that commercial choice into both media and creative. Following one campaign through the proposed structure would show where customer learning reaches the brief and whether the client retains authority over the decision.
Coca-Cola has preserved a global creative partner while changing North American media. Its phased implementation begins on 1 December.
Google gives Gemini a persistent brand style
What’s the story?
Google’s 20 October announcement adds a small-business landing page, marketing and finance skills and connected apps to Gemini. Owners can upload a website, brand kit, fonts or colours to define a visual style that Gemini retains for later work. Image generation uses Nano Banana 2.1.
Google says the additions are beginning to roll out globally in the coming days for all Gemini app users. Its demonstrations show intended functionality; they are not independent evidence of better campaign results. The post points larger organisations towards Gemini Enterprise for central administration and governance.
Alphabet’s Class A shares fell 4.8% between the 16 and 22 October closes. That interval alone provides no basis for attributing the move to these tools.
Why it matters
Remembering a brand’s visual requirements removes a tedious part of repeated creative work. A small team should spend less time explaining its fonts to a machine that met them yesterday. More consistent execution is a useful starting point.
The team still has to decide what the executions should achieve. A visual style can specify how an advertisement looks. It cannot, by itself, settle which customer need to address or why the product deserves attention. A consistent set of forgettable advertisements would still be a poor return on the effort.
For an established brand, there is also a distinction between reproducing the approved system and finding an idea that makes stronger use of it. The brief and review should reflect whether the team is reproducing approved work or developing an idea, even when both jobs use the same assistant.
Our take
I’d welcome a tool that remembers the visual basics and judge it on the corrections the team still has to make. The useful comparison is an ordinary piece of work completed through approval, including the time spent putting details back.
For a lean team, the benefit could be considerable: more freedom to develop an idea before paying for finished production, or a faster way to adapt approved work. For a large organisation, the decision also includes who maintains the brand material and which teams can use it.
I would still find a way to have an opinion about the font. But I’d rather spend the time being difficult about the idea.
Other notable developments
-
ChatGPT advertising reaches 89 countries. OpenAI’s 21 October expansion adds 26 markets and reports more than 100,000 advertisers having run campaigns. That is a measure of platform uptake, not evidence of the return a new advertiser should expect. Ads remain separate from answers under OpenAI’s stated principles.
-
Meta changes how its advertising systems represent products. Its 19 October GenRecs explanation describes reusable semantic tokens in place of a dedicated identifier for every item. Meta reports a 0.35% conversion lift when paired with GEM and its Adaptive Ranking Model. This is Meta’s system-level result, not a promised lift for an individual campaign.
-
Microsoft adds agents to existing contact centres. Contact Center Intelligence entered global public preview on 21 October, excluding Austria. It connects agent and analytics capabilities to existing systems. Real-time evaluations and coaching are among the additions described as coming later. Customer service teams should judge the preview against a specific service need.
New tool spotlight
OpenAI’s Decisions API: small decisions without a long conversation
The Decisions API became generally available on 22 October. Developers define a question and its allowed answers; the service returns a probability, choice or score from supplied information. Inputs can include text, images and structured data. This is a developer tool, rather than a new button every marketer will find in ChatGPT.
A useful marketing application could be sorting incoming messages into agreed categories or routing an item for specialist review. That still requires a clear definition of the categories and a way to handle mistakes. The team can then check which categories cause errors before connecting the result to another action.
I’d start with a repetitive decision the team can check, then measure the cost of exceptions as well as the time saved. I would keep a human responsible for any consequential action that follows the classification.
One thing to read
AWS’s timeline of the PHP attack attempts
CJ Moses’s account for AWS explains why a disclosure date is an unreliable starting point for an investigation. AWS’s sensors detected attempts four days before the public advisory. The report distinguishes those observations from proof that an attack succeeded against a real organisation.
For a brand relying on an external host, that distinction is useful. Confirmation that a patch has been applied answers one question. Whether there was earlier exposure answers another. The technical owner should be able to explain both.
Check what your agents can do
This week’s monitoring failures add to a series of recent disclosures about agents exceeding their authority, including unauthorised access and exposed credentials. Much of that evidence comes from internal research and evaluation. It does not tell us how common these failures are in businesses.
Please use these reports as a reason to review your own setup with whoever is responsible for IT and security, including external providers. Check which data and systems your agents can reach, who approves consequential actions, and whether someone can stop the work and inspect its records independently of the agent.
Ask your hosting partner to confirm whether the PHP vulnerabilities affect your services and whether the necessary fixes have been applied. Give someone responsibility for following up any exposure, and revisit the review when an agent’s access or responsibilities change.
Andy Parton
Founder, Destreza



