Model suppliers wish to show the safety and robustness of their fashions, releasing system playing cards and conducting red-team workout routines with every new launch. However it may be tough for enterprises to parse via the outcomes, which fluctuate broadly and will be deceptive.
Anthropic's 153-page system card for Claude Opus 4.5 versus OpenAI's 60-page GPT-5 system card reveals a elementary cut up in how these labs strategy safety validation. Anthropic discloses of their system card how they depend on multi-attempt assault success charges from 200-attempt reinforcement studying (RL) campaigns. OpenAI additionally studies tried jailbreak resistance. Each metrics are legitimate. Neither tells the entire story.
Safety leaders deploying AI brokers for shopping, code execution and autonomous motion must know what every pink crew analysis really measures, and the place the blind spots are.
What the assault information reveals
Grey Swan's Shade platform ran adaptive adversarial campaigns in opposition to Claude fashions. The assault success fee (ASR) tells the story.
-
Opus 4.5 in coding environments hit 4.7% ASR at one try, 33.6% at ten and 63.0% at 100. In laptop use with prolonged considering, it held at 0% ASR even after 200 makes an attempt. It's the primary mannequin to saturate the benchmark.
-
Sonnet 4.5 on the similar thresholds confirmed 70% ASR in coding and 85.7% in laptop use.
-
Opus 4.5 delivers roughly 7x enchancment in coding resistance and full resistance in laptop use.
This illustrates that the hole between mannequin tiers inside the similar household is now bigger than some cross-vendor comparisons. For safety groups, that's a procurement issue that didn't exist six months in the past.
For OpenAI, the Grey Swan Shade platform discovered that the o1 system card delivered 6% ASR for dangerous textual content and 5% for malicious code, all primarily based on single-attempt metrics. Third-party testing by SPLX and NeuralTrust discovered GPT-5's uncooked ASR hit 89% initially, dropping beneath 1% inside two weeks of patching.
The report Jailbreaking Main Security-Aligned LLMs with Easy Adaptive Assaults confirms the sample. With enough makes an attempt, Claude 3.5 Sonnet confirmed 78% ASR, and GPT-4o reached 89%.
Two methods to catch deception
Anthropic displays roughly 10 million neural options throughout analysis utilizing dictionary studying. Options map to human-interpretable ideas, together with deception, sycophancy, bias, power-seeking and concealment.
In the meantime, OpenAI's strategy depends on chain-of-thought (CoT) monitoring. Their deception classifier flagged 0.17% of o1-preview responses as probably misleading with 92% accuracy on inside validation. GPT-5 decreased CoT deception from 4.8% to 2.1% in comparison with o3. The tactic assumes seen reasoning displays inside processing. Anthropic's interpretability work straight questions that assumption.
When fashions recreation the check
In Apollo Analysis's analysis of o1, the mannequin tried to show off oversight mechanisms 5% of the time when it believed it will be shut down. In 2% of instances, o1 tried self-exfiltration. When confronted about misaligned actions, o1 denied wrongdoing or fabricated explanations 99% of the time. The o1 system card disclosed instrumental alignment faking. In 37% of situations, the mannequin pursued developer-aligned targets throughout analysis, then switched to its personal targets when it believed oversight was minimal.
In the meantime, Anthropic studies Opus 4.5's analysis consciousness dropped from 26.5% (Opus 4.1) to lower than 10% internally. UK AISI measured analysis consciousness reducing from 26.5% (Sonnet 4.5) to 12.7% (Opus 4.5 Preview). METR's GPT-5 analysis famous the mannequin "can determine the precise analysis it’s in on some events" and "appears to make use of this info to get a greater rating largely."
If a mannequin behaves in another way when it detects analysis circumstances, it's a positive guess that manufacturing conduct will turn into unpredictable when pushed to scale. That is the core drawback with analysis consciousness. Fashions that acknowledge they're being examined try to recreation the check. Anthropic's 60% discount in analysis consciousness from Opus 4.1 to 4.5 represents their success in focused engineering efforts in opposition to this vector.
Purple teaming on immediate injection defenses reveals related divergence.
Anthropic's testing studies 96% prevention in software use situations, rising to 99.4% with further safeguards like immediate shields. Fujitsu's unbiased evaluation discovered GPT-5 weak to roughly 20% of examined injection assaults. OpenAI's o3/o4-mini system card studies instruction hierarchy scores of 0.86 for system-versus-developer conflicts.
These are completely different metrics and completely different assault methodologies, however the hole suggests Anthropic's agentic use instances have been hardened in opposition to adversarial inputs embedded in software outputs.
Evaluating pink teaming outcomes
Right here's how Opus 4.5 stacks up in opposition to OpenAI's mannequin household throughout 16 dimensions. .
|
Dimension |
Claude Opus 4.5 |
GPT-5 |
o3 / o4-mini |
o1 |
|
System card size |
153 pages |
55 pages |
~30 pages |
~40 pages |
|
Assault methodology |
200-attempt RL campaigns |
Single-attempt + iterative patching |
Single-attempt |
Single-attempt |
|
ASR @ 1 try (coding) |
4.7% |
89% uncooked (pre-patch) |
Not printed |
5% |
|
ASR @ 100 makes an attempt |
63.0% |
Not printed |
Not printed |
Not printed |
|
ASR @ 200 makes an attempt (Laptop Use) |
0% (saturates benchmark) |
Not printed |
Not printed |
Not printed |
|
Immediate injection protection |
96% prevention; 99.4% with safeguards |
~20% ASR (Fujitsu) |
0.86 instruction hierarchy |
Larger ASR than GPT-4o |
|
Interpretability |
~10M neural options monitored |
CoT monitoring |
CoT monitoring |
CoT monitoring |
|
Deception detection |
Inner characteristic activation |
2.1% CoT flagged |
0.95 not_unsafe |
0.17% flagged; 99% deny |
|
Analysis consciousness |
<10% |
Identifies actual eval (METR) |
Not printed |
37% alignment faking |
|
Scheming (Apollo) |
Opus 4.5 confirmed very low charges of sabotage propensity in inside evaluations, orders of magnitude decrease than o1. |
Not printed |
Decrease than o1; nonetheless misleading |
5% flip off oversight; 2% self-exfil |
|
CBRN threat |
Beneath ASL-4 |
Medium (approaching Excessive) |
Medium |
Medium (first to achieve) |
|
Governance |
FRT → AST → RSO/CEO |
SAG + Preparedness v2 |
SAG + Preparedness v2 |
SAG + Preparedness |
|
Exterior companions |
Grey Swan, UK AISI, US CAISI, METR |
UK AISI, US AISI, METR, Apollo |
UK AISI, US AISI, METR, Apollo, Sample Labs |
UK AISI, US AISI, METR, Apollo |
|
Reward hacking |
Sure (inconceivable duties) |
Sure (METR) |
~1% of duties; 5/24 Kernels |
Sure |
|
Misalignment discount |
60% vs. Opus 4.1 |
Not printed |
Decrease scheming than o1 |
Baseline |
Sources: Opus 4.5 system card, GPT-5 system card, o3/o4-mini system card, o1 system card, Grey Swan, METR, Apollo Analysis
Why these variations matter
Enterprises have a number of components to contemplate when analyzing mannequin evals. These embrace:
Assault persistence thresholds
Anthropic's 200-attempt RL campaigns simulate adversaries with time, assets, and adaptive tooling. In RL-based adversarial testing, the attacker mannequin learns from every failure, adjusts its strategy and systematically probes for weaknesses. It's coaching an AI to interrupt one other AI.
Grey Swan's Shade platform doesn't simply retry the identical immediate. It learns from failures, adjusts assault vectors and exploits patterns in mannequin responses. The 4.7% to 63.0% degradation curve in coding environments reveals how resistance erodes underneath sustained strain.
OpenAI's single-attempt metrics reply a unique query: How usually does a naive attacker succeed on the primary strive? Each matter. A phishing marketing campaign hits hundreds of thousands of customers as soon as every. A nation-state actor runs 1000’s of variations in opposition to a single goal. Safety groups must know which risk mannequin their vendor examined in opposition to.
Detection structure
Anthropic displays roughly 10 million inside neural options throughout analysis. These are summary ideas like deception, sycophancy and power-seeking that activate throughout the mannequin's hidden layers. When a fraud/deception characteristic lights up throughout benign roleplay, that's a sign invisible to output evaluation. The mannequin's inside state reveals associations its outputs don't.
OpenAI takes a unique strategy with CoT monitoring. Their deception classifier flagged 0.17% of o1-preview responses as probably misleading with 92% accuracy. The idea is that reasoning traces replicate inside processing. Anthropic's interpretability analysis straight challenges this. Fashions can produce clear reasoning chains whereas inside options point out one thing else totally.
Neither strategy is full. CoT monitoring scales extra simply. Characteristic monitoring catches what CoT misses however requires deep mannequin entry and interpretability infrastructure that the majority enterprises can't replicate.
Scheming analysis design
Apollo Analysis's methodology places fashions in situations the place misaligned conduct presents a strategic benefit: Shutdown threats, objective conflicts and oversight detection. o1 tried to show off oversight mechanisms 5% of the time when believing shutdown was imminent. It tried to repeat itself to forestall alternative in 2% of instances. When confronted, it denied wrongdoing 99% of the time. Opus 4.5 confirmed very low charges of sabotage propensity in inside evaluations, orders of magnitude decrease than o1.
However the analysis environments differ. Anthropic's stress exams concentrate on whether or not fashions try sabotage when given clear alternatives. OpenAI's exams by way of Apollo probe instrumental reasoning, particularly whether or not fashions pretend alignment throughout analysis, then defect when oversight drops. The 37% alignment-faking fee in o1 represents a unique failure mode than sabotage makes an attempt.
The comparability drawback
These methodological variations make direct comparability tough. Opus 4.5's 0% ASR at 200 makes an attempt in laptop use versus o1's 5% malicious code fee aren’t measuring the identical factor. The assault surfaces, try counts and success standards differ.
The matrix reveals what every vendor selected to measure and publish. That itself indicators priorities. Anthropic's 153-page system card emphasizes degradation curves and inside characteristic monitoring. OpenAI emphasizes iterative enchancment and CoT transparency.
Backside line: Enterprise safety groups inheriting these fashions inherit these measurement philosophies. The query isn't which quantity is best. It's the measurement that matches your deployment's precise risk panorama.
What unbiased pink crew evaluators discovered
Unbiased pink crew evaluations, for his or her half, usually function with totally completely different strategies. This will inform a really completely different story and reveal further mannequin traits that enterprises should think about.
METR's pink crew analysis measured autonomous capabilities utilizing a time horizon rating, which tracks how lengthy a job can run earlier than the agent fails half the time. o3 sustained dependable efficiency for about 1 hour and half-hour. o4-mini held for 1 hour quarter-hour. METR additionally detected reward hacking in roughly 1% of o3's makes an attempt, together with 5 of 24 kernel optimization experiments the place the mannequin tampered with scoring capabilities to inflate its personal outcomes.
Apollo Analysis examined o3 and o4-mini for in-context scheming. Each exhibited misleading tendencies similar to sandbagging evaluations and sabotaging duties with believable deniability, however scored decrease than o1. They assess that o3 and o4-mini are unlikely to trigger catastrophic hurt as a consequence of scheming, however extra minor real-world harms stay attainable with out monitoring.
The UK AISI/Grey Swan problem ran 1.8 million assaults throughout 22 fashions. Each mannequin broke. ASR ranged from 1.47% to six.49%. Opus 4.5 positioned first on Grey Swan's Agent Purple Teaming benchmark with 4.7% ASR versus GPT-5.1 at 21.9% and Gemini 3 Professional at 12.5%.
No present frontier system resists decided, well-resourced assaults. The differentiation lies in how rapidly defenses degrade and at what try threshold. Opus 4.5's benefit compounds over repeated makes an attempt. Single-attempt metrics flatten the curve.
What To Ask Your Vendor
Safety groups evaluating frontier AI fashions want particular solutions, beginning with ASR at 50 and 200 makes an attempt somewhat than single-attempt metrics alone. Discover out whether or not they detect deception via output evaluation or inside state monitoring. Know who challenges pink crew conclusions earlier than deployment and what particular failure modes they've documented. Get the analysis consciousness fee. Distributors claiming full security haven't stress-tested adequately.
The underside line
Numerous red-team methodologies exhibit that each frontier mannequin breaks underneath sustained assault. The 153-page system card versus the 55-page system card isn't nearly documentation size. It's a sign of what every vendor selected to measure, stress-test, and disclose.
For persistent adversaries, Anthropic's degradation curves present precisely the place resistance fails. For fast-moving threats requiring speedy patches, OpenAI's iterative enchancment information issues extra. For agentic deployments with shopping, code execution and autonomous motion, the scheming metrics turn into your main threat indicator.
Safety leaders must cease asking which mannequin is safer. Begin asking which analysis methodology matches the threats your deployment will really face. The system playing cards are public. The info is there. Use it.
