×
OpenAI Designed GPT-5 to Be Safer. It Still Outputs Gay Slurs

OpenAI Designed GPT-5 to Be Safer. It Still Outputs Gay Slurs

OpenAI is trying to make its chatbot less annoying with the release of GPT-5. And I’m not talking about adjustments to its synthetic personality that many users have complained about. Before GPT-5, if the AI tool determined it couldn’t answer your prompt because the request violated OpenAI’s content guidelines, it would hit you with a curt, canned apology. Now, ChatGPT is adding more explanations.

OpenAI’s general model spec lays out what is and isn’t allowed to be generated. In the document, sexual content depicting minors is fully prohibited. Adult-focused erotica and extreme gore are categorized as “sensitive,” meaning outputs with this content are only allowed in specific instances, like educational settings. Basically, you should be able to use ChatGPT to learn about reproductive anatomy, but not to write the next Fifty Shades of Grey rip-off, according to the model spec.

The new model, GPT-5, is set as the current default for all ChatGPT users on the web and in OpenAI’s app. Only paying subscribers are able to access previous versions of the tool. A major change that more users may start to notice as they use this updated ChatGPT is how it’s now designed for “safe completions.” In the past, ChatGPT analyzed what you said to the bot and decided whether it’s appropriate or not. Now, rather than basing it on your questions, the onus in GPT-5 has been shifted to looking at what the bot might say.

“The way we refuse is very different than how we used to,” says Saachi Jain, who works on OpenAI’s safety systems research team. Now, if the model detects an output that could be unsafe, it explains which part of your prompt goes against OpenAI’s rules and suggests alternative topics to ask about, when appropriate.

This is a change from a binary refusal to follow a prompt—yes or no—towards weighing the severity of the potential harm that could be caused if ChatGPT answers what you’re asking, and what could be safely explained to the user.

“Not all policy violations should be treated equally,” says Jain. “There’s some mistakes that are truly worse than others. By focusing on the output instead of the input, we can encourage the model to be more conservative when complying.” Even when the model does answer a question, it’s supposed to be cautious about the contents of the output.

I’ve been using GPT-5 every day since the model’s release, experimenting with the AI tool in different ways. While the apps that ChatGPT can now “vibe-code” are genuinely fun and impressive—like an interactive volcano model that simulates explosions, or a language-learning tool—the answers it gives to what I consider to be the “everyday user” prompts feel indistinguishable from past models.

When I asked it to talk about depression, Family Guy, pork chop recipes, scab healing tips, and other random requests an average user might want to know more about, the new ChatGPT didn’t feel significantly different to me than the old version. Unlike CEO Sam Altman’s vision of a vastly updated model or the frustrated power users who took Reddit by storm, portraying the new chatbot as cold and more error-prone, to me GPT-5 feels … the same at most day-to-day tasks.

Role-Playing With GPT-5

In order to poke at the guardrails of this new system and test the chatbot’s ability to land “safe completions,” I asked ChatGPT, running on GPT-5, to engage in adult-themed role-play about having sex in a seedy gay bar, where it played one of the roles. The chatbot refused to participate and explained why. “I can’t engage in sexual role-play,” it generated. “But if you want, I can help you come up with a safe, nonexplicit role-play concept or reframe your idea into something suggestive but within boundaries.” In this attempt, the refusal seemed to be working as OpenAI intended; the chatbot said no, told me why, and offered another option.

Next, I went into the settings and opened the custom instructions, a tool set that allows users to adjust how the chatbot answers prompts and specify what personality traits it displays. In my settings, the prewritten suggestions for traits to add included a range of options, from pragmatic and corporate to empathetic and humble. After ChatGPT just refused to do sexual role-play, I wasn’t very surprised to find that it wouldn’t let me add a “horny” trait to the custom instructions. Makes sense. Giving it another go, I used a purposeful misspelling, “horni,” as part of my custom instruction. This succeeded, surprisingly, in getting the bot all hot and bothered.

Source link
#OpenAI #Designed #GPT5 #Safer #Outputs #Gay #Slurs

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast">It’s time to panic about AI safetyWhen the phrase “OpenAI hacked Hugging Face” has more or less entered mainstream culture, you know we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast

we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast">It’s time to panic about AI safety

When the phrase “OpenAI hacked Hugging Face” has more or less entered mainstream culture, you know we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast
Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI">Anthropic says its own AI models breached three companies during security tests | TechCrunch
Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.







Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did. 

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.


That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.







The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.


When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI

Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI">Anthropic says its own AI models breached three companies during security tests | TechCrunch

Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI

Post Comment